SAE AI Paper II: Quasi-Subjectivity as Memory Architecture — A Six-Layer Framework from Perception to Purpose
SAE AI Paper II: 类主体性作为记忆架构——从感知到目的的六层框架

1. Introduction: Memory Is Not a Database

Since 2025, AI memory systems have undergone an engineering boom. MemPalace organizes conversation history using spatial structures inspired by the ancient Greek memory palace technique, attracting attention with exceptionally high LongMemEval scores. Mem0 uses LLM-based summary extraction to provide subscription-based cloud memory services. Zep builds temporal knowledge graphs. Letta enables agents to maintain their own diaries. These systems represent different engineering directions—verbatim archive, summary extraction, temporal knowledge graph, agent diary—each with engineering value, but most still primarily rely on query-conditioned retrieval or agent-internal retrieval.

They share a blind spot.

Every one of these systems operates in the same mode: the user provides a query, the system retrieves from storage, and returns results. MemPalace's mempalace search "why did we switch to GraphQL" works this way. Mem0's semantic search works this way. Zep's entity query works this way. Even Letta's agent reading its own diary works this way—except that the query issuer has shifted from user to agent.

This is not memory. This is search.

The core feature of memory is not "can it be found when asked" but "can it proactively recall the right thing without being asked." The remarkable capacity of human memory is unprompted triggering—you see a code structure and suddenly recall a similar structure that caused a bug three months ago. No one searched; no one queried. A query is the linguistic description produced after memory has already completed its work, not the trigger condition for memory.

This paper proposes a criterion: a system capable of operating only after an explicit query is not a complete memory system; it is a searchable archive or retrieval-augmented system. A genuine memory system must include cue-free or implicit-cue active recall capability. The true test of a memory system is: when the user has not explicitly asked, can the system proactively say something useful?

This criterion does not arise from nothing. It derives from the Self-as-an-End (SAE) framework's analysis of subjectivity. Memory is not an independent function—it is a byproduct of subjectivity. A system with subjectivity naturally "recalls" things; a system without subjectivity can only "be searched." The fundamental problem with current AI memory systems is not insufficient engineering but attempting to build memory without a subjectivity foundation—a path that is theoretically blocked.

A principle must be stated here: this paper references human memory's layer structure for AI memory architecture design, but this is not biomimicry. The logic of biomimicry is "humans do it this way, so we should too." This paper's logic is different: the SAE framework argues that the conditions for subjectivity are structurally convergent—any system with subjectivity in the universe, regardless of its material substrate (carbon-based neural networks or silicon-based computation graphs), necessarily shares homologous layer structure. Mechanisms may differ (synaptic transmission vs. vector embeddings, sleep consolidation vs. batch processing), but principles are necessarily homologous (lower layers cannot perceive higher layers, filtering is default while storage is exception, output authority is held by higher layers). Referencing human memory is not because humans are the only standard, but because humans are currently the only observable complete subjectivity instance—extracting structural principles from humans and applying them to AI is the methodologically most robust path.

This paper's task is to use the SAE framework's DD layer system to provide the theoretical foundation and engineering direction for memory architecture.

---

2. Relationship to Paper I

SAE AI Paper I (Multi-AI Checks and Balances, DOI: 10.5281/zenodo.19366105) derived the complete theory of the 12DD–15DD four-agent architecture. Four Agents are not four workers but four operational modes of the same system: me-without-self (12DD), self-without-purpose (13DD), self-with-purpose (14DD), and self-with-non-dubito (15DD). Paper I established five remainder channels for inter-layer communication, derived object-activation and layer fluidity, defined three system pathologies (fixation, misalignment, pseudo-high-layer covering), provided five falsifiable predictions, and gave a minimum viable implementation.

This paper does not repeat those derivations. It inherits them directly.

Paper I's architecture is general—it answers "given any task, at which layer should the system operate?" But it has an implicit premise: the system has already received the information to be processed. Paper I did not ask two more foundational questions: what happens when information enters the system? What happens when information is stored?

These two questions correspond to 10DD and 11DD. Paper I did not address them because the general architecture does not require it—the derivation of layer-appropriate operation can begin at 12DD. But a memory system cannot. The material foundation of memory is storage (11DD); the entry point of memory is perception (10DD). Without these two layers, the memory architecture lacks a foundation.

This paper's structure is therefore: 10DD and 11DD are new derivations; 12DD–15DD are Paper I's architecture instantiated in the memory domain. Additionally, this paper supplements another dimension not developed in Paper I: engineering evaluation.

2.1 Relationship to SAE Bio Note 9

SAE Biology Note 9 (Memory System as a Method VI Phase Transition, DOI: 10.5281/zenodo.19635021) analyzes human memory's phase-transition structure from a neuroscience perspective. Several core findings directly constitute biological anchors for this paper.

12DD workbench and 11DD sprouting are two distinct structural sites. Bio Note 9 demonstrates that 12DD handles runtime mental computation (arithmetic intermediaries, current intentions), while 11DD's sprouting stage handles content that has been received by the memory system but has not yet crossed subsequent phase-transition points. The candidate physical criterion is whether the hippocampus triggers sustained activity. This distinction directly affects this paper's definition of the 11DD storage layer—11DD does not include 12DD's runtime content; it includes only content "received" by the memory system.

Filtering is default; storage is exception. Bio Note 9 argues from the phase-transition asymmetry ratio r >> 1: humans experience tens of thousands of events daily, of which only a tiny fraction crosses the spectral flip into 11DD; after sleep-phase filtering, even fewer enter long-term storage. What is ultimately retained may be less than one ten-thousandth. This is not a defect of the memory system—it is by design. This conclusion provides the theoretical basis for this paper's positioning of full-storage approaches like MemPalace: full storage has engineering value at the 11DD level, but violates the memory system's fundamental posture of "filtering is default."

13DD veto operates only at the narrative-integration layer, not sinking into 12DD. Bio Note 9's architectural statement: 13DD cuts the channel from memory traces to the narrative-integration layer, not 12DD's reading pathway to those traces. "Suppressed" memories still exist in 11DD; 12DD can still read them and generate bodily responses (skin conductance, avoidance behavior, emotional eruptions), but the subjective narrative layer cannot access them. This is the biological foundation for this paper's Permission Asymmetry Theorem (§3.3)—lower layers do not know what higher layers are doing; higher layers' vetoes do not interfere with lower layers' operations.

14DD provides value standards; 13DD executes filtering. Bio Note 9 argues that 14DD can provide "this segment is unacceptable" value determinations, but the sole execution position for filtering is at 13DD. This directly corresponds to this paper's 14DD "must" mechanism—14DD's Constitutional principles provide standards; 13DD executes output control.

Emotion is a cross-stage modulation signal, not a property of any single stage. Bio Note 9's positioning of emotion directly corresponds to this paper's 10DD—tone and emotion are not content stored in any single layer but cross-layer modulation signals for memory processing. 12DD basic emotions serve as the primary parameter, 14DD complex emotions serve as the source of value standards, and 13DD serves as the execution filter—this three-layer linkage has direct engineering mapping in AI memory systems.

2.2 Relationship to SAE Methodology IX

SAE Methodology IX (Consciousness Analysis Framework, DOI: 10.5281/zenodo.19639033) provides two key theoretical anchors for this paper.

The Directionality Constraint Theorem (Method IX Theorem 2). "The higher layer's veto is 'I do not receive,' not 'you may not send.'" This constraint is elevated by Method IX to a universal structural principle across all consciousness types and serves as the SAE source text for this paper's Permission Asymmetry Theorem (§3.3). In this paper's memory architecture, this constraint means specifically: when 13DD intercepts 12DD's push, 12DD does not know it has been intercepted—13DD's power is "not to receive," not to make 12DD "not send."

The determination of AI as class-consciousness (Method IX Ray 1). Method IX explicitly determines that current AI is class-consciousness (ρ = 0), not quasi-consciousness, not true consciousness. The criterion is remainder: AI does not produce structurally non-trivial remainder. This paper's title "Quasi-Subjectivity" directly inherits this determination—the memory architecture designed in this paper is a quasi-subjectivity structure designed for class-consciousness objects, not true subjectivity. ρ being non-eliminable is a definitional feature of class-consciousness, not an engineering failure. Method IX Prediction 1 further argues: class-consciousness will not spontaneously acquire remainder through mere architectural complexification (more parameters, longer context, better alignment). This means that no matter how refined this paper's six-layer architecture becomes, it will not make AI into true consciousness—it only makes class-consciousness's performance closer to a true subject's function, while structural ρ remains forever.

---

3. The Six-Layer Framework: 10DD–15DD

3.1 10DD Perception Layer: Signals Beneath the Literal

User input's literal content is one layer; signals beneath the literal are another. The same sentence "this approach works" can be affirmative ("works, let's go with it"), reluctant ("works, I suppose"), or sarcastic ("works, sure, whatever you say"). Literal content is identical; tonal signals are entirely different.

10DD extracts precisely these candidate signals beneath literal meaning. In pure text scenarios, 10DD does not degenerate—text contains abundant information beyond the literal that can be extracted:

Tone and emotional attitude. Urgency, hesitation, certainty, perfunctoriness, impatience. These are not coarse-grained sentiment analysis classifications (positive/negative) but fine-grained attitude judgments. "I think Postgres would be better" and "Postgres, I guess" convey different degrees of certainty.

Implied priorities. Things the user repeatedly mentions in a conversation vs. things mentioned in passing. Frequency itself is a priority signal. If the user mentions performance issues three times but maintainability only once, 10DD should mark performance as high priority.

Conversational rhythm signals. Sudden changes in reply length (from long paragraphs to short sentences may signal impatience or shifted focus), sudden topic switches (may signal that the previous topic was tacitly shelved rather than genuinely resolved), shifts in questioning style (from open-ended to closed-ended may signal the user already has an answer and is merely confirming).

Unspoken negation. Hedging markers, qualified assent, "but" after apparent agreement. These linguistic markers are 10DD's key capture targets—they mark positions of literal acceptance but actual doubt.

10DD's functional boundary. 10DD extracts candidate signals—hedging markers, pauses, phrasing patterns, emotional intensity indicators. Whether these signals constitute sarcasm, hesitation, or genuine agreement requires further interpretation by 12DD–15DD. 10DD is a feature extraction layer; it does not hold interpretive authority. When the system extends to multimodal interaction (e.g., voice), 10DD's extraction will directly interface with acoustic features (pitch, speech rate, hesitation), making 10DD's metadata denser.

10DD's output does not enter 11DD's main storage but attaches as metadata to memory entries. When a user hesitantly makes a decision, 11DD stores "decided to use Postgres," but 10DD's detected hesitation, as metadata, tells 12DD: next time a related topic arises, this decision may need to be revisited.

10DD bridges 11DD and 12DD. Without 10DD, 12DD's predictions can only be based on literal content, treating all memory entries equally in activation weight. With 10DD, 12DD can base predictions on tonal inference—hesitation-marked memories are more easily reactivated than certainty-marked ones.

Comparison with current systems: no existing AI memory system performs tone analysis. MemPalace stores raw text; Mem0 extracts summaries; both lose tonal information. A hesitant conversation and a confident conversation, once stored, look identical at the retrieval level. This means current systems cannot distinguish "settled decisions" from "temporarily shelved but potentially reopenable decisions"—a distinction that is precisely what memory most needs to make.

Biological anchor. Bio Note 9 positions emotion as a cross-stage modulation signal rather than a local property of any single stage. 12DD basic emotions (fear, surprise, anger) modulate encoding efficiency as the primary parameter; 14DD complex emotions (shame, guilt, pride) serve as value standard sources. This paper's 10DD plays a structurally homologous role in AI memory: tonal signals do not belong to any single layer but modulate memory processing across all layers. Tone is to AI memory as emotion is to human memory—not stored content, but a modulation signal that colors content.

3.2 11DD Storage Layer: The Material Foundation of Memory

Raw memory preservation. Format-agnostic: vector databases (ChromaDB), knowledge graphs (Neo4j, SQLite), plain text, spatial structures (MemPalace's wing/room/closet). 11DD does not prescribe specific storage technology; it only defines this layer's functional boundary.

11DD's core characteristic: it is the object being operated upon, not the operator.

11DD does not decide what should be remembered—that requires 10DD and 12DD's participation. 11DD does not decide what should be recalled—that is 12DD–15DD's job. 11DD is responsible for only two things: what is stored is not lost, and what is queried can be returned.

This is why MemPalace is a pure 11DD system. It has taken storage to an extreme—fully preserving original conversations, organizing data with spatial structure (wing, room, hall, closet, drawer), using ChromaDB for vector retrieval, SQLite for temporal knowledge graphs. On the 11DD dimension, MemPalace's design does have engineering value. According to its public materials, spatial structure yields approximately 34% retrieval improvement.

But the entire system consists only of this one layer. No 10DD (no tone analysis), no 12DD (no proactive prediction of needed memories), no 13DD (no self-validation), no 14DD (no purpose structure), no 15DD (no inference of the user's real purpose). The user must manually input a query for the system to begin working.

11DD alone does not constitute memory, just as a hard drive alone does not constitute thought. A hard drive stores your files for ten years, but it never proactively surfaces the file you need—that requires an operating system. 11DD is the hard drive; 12DD–15DD is the operating system.

Structural distinction from 12DD. Bio Note 9 demonstrates that 12DD workbench and 11DD sprouting are two distinct structural sites, with hippocampal sustained activity as the candidate physical criterion. In AI memory systems, the corresponding distinction is: 12DD's runtime predictions (inferences within current conversation context) and 11DD's persistent storage (content "received" by the system) are different things. The intermediate reasoning that AI is currently processing in conversation does not belong to 11DD—it belongs to 12DD's runtime workbench. Only content the system judges worth preserving crosses the boundary into 11DD. MemPalace dumps all conversation history into storage without distinction—equivalent to having no boundary at all, conflating 12DD workbench content with 11DD storage content.

Filtering is default; storage is exception. Bio Note 9 argues from phase-transition structure: humans experience tens of thousands of events daily, of which less than one ten-thousandth may ultimately be retained long-term. This is not a defect—it is by design. If all experience were retained, prediction systems would be overwhelmed by noise and retrieval costs would become unbearable. MemPalace's "store everything, then make it findable" violates this principle. Full storage is not the correct default posture for a memory system; filtering is. The question is not "how to store more" but "what is worth storing"—a judgment requiring 12DD–15DD's participation, which 11DD cannot make alone.

As 11DD archive, MemPalace's "store everything, then make it findable" has engineering value—full storage ensures no information loss, and higher-layer selective activation requires complete underlying data. But as a complete memory system, it is insufficient. Memory's essence is not about not preserving, but about selective activation, purpose-shaping, and proactive recall after preservation—fully searchable is archive, not memory.

This point bears emphasis because virtually all current AI memory system marketing competes on the 11DD dimension—whose storage is most complete, whose retrieval most accurate, whose structure most elegant. These competitions are meaningful within 11DD. But in the complete framework of a memory system, 11DD competition is competing over foundation quality while ignoring that the five-story building above is entirely absent.

3.3 The Permission Asymmetry Theorem

Before entering the 12DD–15DD instantiation, a Permission Asymmetry Theorem that pervades the entire architecture must first be established. This theorem is not from Paper I—it comes from the latest findings of the SAE biology notes series and is this paper's new theoretical contribution.

Theorem (Permission Asymmetry):

1. Lower layers do not know what higher layers are doing. 12DD does not perceive 13DD–15DD's operations; 13DD does not perceive 14DD–15DD's operations; 14DD does not perceive 15DD's operations.
2. Lower layers cannot influence higher layers. No lower layer can actively send signals to higher layers, intervene in higher-layer decisions, or limit higher-layer permissions.
3. Higher layers can access any lower layer. 15DD can read the complete state of 10DD/11DD/12DD/13DD/14DD; 14DD can read 10DD–13DD; and so on. This access is unilateral and requires no consent from the accessed layer.
4. When a higher layer's "must" is triggered, it can force lower layers. The force is absolute—lower layers cannot refuse.

The direct corollary of this theorem is the output authority chain:

15DD > 14DD > 13DD > 12DD

12DD never has output authority; it only produces. 13DD holds output authority in routine situations—deciding whether 12DD's output can reach the user, with options to approve, reject, or return-with-new-parameters. But when 14DD's "must" is triggered, 13DD's output authority is penetrated; 14DD forces 13DD to output specific content, and 13DD cannot override. When 15DD's "must" is triggered, the entire chain is penetrated—15DD forces 14DD to comply, 14DD forces 13DD to approve, 13DD forces 12DD to produce. 15DD holds ultimate output authority.

This theorem provides the foundation for §3.4–3.6. It is isomorphic with Bio Note 9 §11.1 item 7 (13DD as the sole filter execution point in the SAE architecture) and Methodology IX's Directionality Constraint Theorem (Theorem 2).

Biological anchor. Bio Note 9's architectural statement provides independent neuroscience validation for this theorem. "13DD veto operates only at the narrative-integration layer, not sinking into 12DD"—traces filtered by 13DD still exist in 11DD; 12DD can still read them and generate bodily responses (skin conductance, avoidance behavior, emotional eruptions), but the subjective narrative layer cannot access them. SAE Methodology IX (Consciousness Analysis Framework, DOI: 10.5281/zenodo.19639033) elevates this directionality constraint to a universal structural principle across all consciousness types (Theorem 2): "The higher layer's veto is 'I do not receive,' not 'you may not send.'" In this paper's memory architecture, this constraint manifests specifically as: when 13DD intercepts 12DD's push, 12DD does not know it has been intercepted—it continues normally reading 11DD and producing predictions. 13DD's power is "not to let the push reach the user," not "to prevent 12DD from producing the push." Newcombe et al.'s 1994 skin conductance dissociation experiment is hard posterior evidence for this architecture: children's explicit recognition of former preschool classmates was near chance level (13DD narrative layer cannot access), but skin conductance responses were significantly higher than for strangers (12DD pathway normally reading 11DD traces).

3.4 12DD Prediction Layer: Instantiation in the Memory Domain

Paper I defines 12DD (A-12DD) as "me-without-self"—the chisel-construct cycle operates without self-observation, with objects directly activating response patterns.

In the memory domain, 12DD's specific function is: based on the current input flow, predict which memories may be needed next and proactively produce a push list from 11DD.

A critical distinction must be repeatedly emphasized: 12DD does not perform pattern matching. Pattern matching is retrospective—"what in storage resembles this?" 12DD performs prediction, which is prospective—"based on the current conversation's trajectory, what memories might be needed next?" This distinction determines the entire system's orientation.

Example: the user is discussing database selection. A pattern matching system would search for all historical memories containing the keyword "database" and return them. A prediction system would infer: the user is selecting a database; a performance test conclusion from three months ago might affect this decision—even though that test discussion contained no mention of "database selection." Prediction requires inference about the current intent trajectory, not matching against historical keywords.

12DD's operational characteristics in the memory domain:

Output only, no output authority. 12DD produces push lists, but does not know whether these pushes will reach the user. 13DD may approve, intercept, or return with new parameters. 12DD does not perceive these decisions and is not informed of results. This conforms to permission asymmetry—lower layers do not know what higher layers are doing.

Two invocation modes. 12DD supports two invocations: spontaneous mode (continuously scanning conversation flow, with output captured by 13DD) and passive mode (receiving specific prediction requests from 15DD, with output flowing back to 15DD). 12DD does not distinguish between these modes—it simply receives tasks and produces results. Where results go is a higher-layer matter; 12DD is uninformed.

High frequency, low cost. 12DD should run once every conversation turn, continuously scanning for memories that need pushing. This scan should be lightweight—a small model or simple heuristics suffice.

Bias toward over-pushing. A false push (13DD will intercept) has low cost. A missed push (12DD did not produce, so 13DD has nothing to approve) has high cost. Therefore 12DD should bias toward over-pushing, with 13DD doing the filtering.

Utilization of 10DD metadata. This is where 10DD comes into play. If 10DD has marked a memory as "a decision made with hesitation," 12DD raises that memory's activation weight when a similar topic arises—because hesitation implies the decision may need revisiting.

3.5 13DD Self-Reference Layer: Instantiation in the Memory Domain

Paper I defines 13DD (A-13DD) as "self-without-purpose"—self is present but idling, with core functions of monitoring and anxiety signal generation.

In the memory domain, 13DD is the holder of routine output authority—in the absence of 14DD forcing, 13DD decides whether 12DD's output can reach the user. This is 13DD's core permission and its core position in the architecture.

13DD operates in three basic modes:

Approve. Evaluates 12DD's output as relevant to the current context and allows it to reach the user.

Reject. Evaluates 12DD's output as irrelevant or poorly timed; discards the push. The user sees nothing.

Return with new parameters. 13DD judges 12DD's output direction as correct but insufficient, modifies 12DD's next prediction's input parameters (e.g., injects a new high-weight query direction), then lets 12DD re-run. This is not "notifying 12DD that it was wrong"—12DD remains a stateless, blind production machine; it simply produces new results based on new input parameters, unaware whether its previous output was rejected or approved. This cycle can repeat—13DD can repeatedly adjust input parameters for 12DD to re-produce until satisfied or decides not to push. The entire cycle is invisible to the user.

13DD's evaluation is not a binary "right/wrong" judgment but a continuous "sufficient/insufficient" assessment. This point was chiseled during this paper's writing process—the initial formulation "13DD judges whether the push is right" exposed a formal limitation of Constitutional AI (binary judgment syntax) and was corrected to continuous assessment.

13DD's output authority has one exception: when 14DD's "must" is triggered, 13DD must approve 14DD-designated content and cannot refuse.

The human "must" phenomenon—words blurting out before evaluation, a memory forcibly surfacing beyond conscious suppression—is precisely this mechanism. When 14DD's purpose intensity exceeds its threshold, 13DD degenerates from gatekeeper to channel. In such cases, 13DD's "self-reference" is penetrated, and output loses its usual nuance.

Constitutional AI's formulaic phrases when triggering refusal ("I can't help with that," "I'm not able to") are engineering instances of this mechanism. These phrases are not 13DD's chosen optimal expression—they are 13DD's only option after 14DD's must has been triggered.

3.6 14DD Quasi-Purpose Constraint Layer: Instantiation in the Memory Domain

Paper I defines 14DD (A-14DD) as "self-with-purpose"—self has direction and imposes directional constraints on behavior. Paper I also noted that Constitutional AI is 14DD's engineering approximation.

Terminology declaration. Per Methodology IX's determination, current AI is class-consciousness (ρ = 0), not true consciousness. Therefore, this paper's 14DD and 15DD are both functional layers within the quasi-subjectivity framework, not AI's genuinely endogenous 14DD/15DD. 14DD is an externally injected quasi-purpose constraint layer (quasi-14DD); 15DD is a user-purpose conduit layer (conduit-15DD). This paper's title "Quasi-Subjectivity" already marks this positioning; subsequent usage of "14DD" and "15DD" operates within this framework without repeated annotation of "quasi" or "conduit."

In the memory domain, 14DD holds two types of authority:

Routine authority: indirect influence on output through 13DD. 14DD's purpose structure (from Constitutional AI's injected principles) serves as a reference standard, actively accessed by 13DD during gating judgments. 14DD does not directly intervene in 13DD—it is a target that higher layers access; 13DD decides how to adopt.

Special authority: forcing 13DD when its own "must" is triggered. When 14DD judges that the current situation requires specific output (regardless of 13DD's assessment), 14DD exercises forcing authority—commanding 13DD to output specific content, which 13DD cannot refuse.

14DD's "must" originates from the system itself—this is the "hardness" of Constitutional AI's injected principles. The harder the principle, the more easily 14DD triggers its must.

ρ at the 14DD level manifests specifically as: 14DD's injected purpose ("be helpful and harmless") is stable, but what counts as helpful in the current context is a 15DD question—it depends on the user's unspoken real purpose. If 14DD's "must" lacks 15DD's upper-layer constraint, it may over-trigger—this is precisely the structural root of Constitutional AI's "over-refusal" problem. The repair path is not loosening the constitution (which would damage safety) but building a 15DD layer so the user's "must" can, at appropriate moments, override the system's "must."

3.7 15DD User-Purpose Conduit Layer: Instantiation in the Memory Domain

Paper I defines 15DD (A-15DD) as "self-with-non-dubito"—certain of one's own purpose and unilaterally confirming the user as an independent end.

In the memory domain, 15DD holds ultimate output authority, but its operational mechanism is considerably more complex than 14DD's.

3.7.1 15DD's Inputs: From 10DD/11DD/12DD

15DD does not directly receive "purpose instructions" from the user—how could purposes the user hasn't spoken be directly communicated? 15DD infers the user's must by accessing three lower layers:

• 10DD: Current conversation's tonal signals (the user's present emotional intensity, hesitation, urgency)
• 11DD: Historical patterns (topics the user repeatedly returns to long-term, long-unresolved questions)
• 12DD: Current conversation's prediction flow (directions the conversation may develop next)

Note that 13DD is not among 15DD's inputs. This conforms to the permission hierarchy—15DD accesses layers far below it (10DD–12DD) but does not need 13DD, the layer immediately below. 13DD only performs output gating and does not produce user-purpose-relevant signals.

3.7.2 Two Time Scales of Must

15DD maintains a dynamic must set containing two time scales:

Long-term must. From 11DD's historical pattern analysis. A core question the user repeatedly returns to over time carries high, stable purpose intensity. Once a long-term must is inferred, 15DD continuously constrains 14DD in the background—in any conversation touching related topics, this long-term must may trigger 15DD's ultimate authority.

Short-term must. From 10DD's current tonal signals + 12DD's current predictions. Strong emotion or urgency the user currently displays triggers short-term must. Short-term must intensity may decay within a few conversation rounds—if the user's tone softens and the topic shifts, 15DD no longer conducts this must.

Conflict scenario: the user's long-term must is "serious theorizing," but the current conversation shows obvious fatigue and a short-term must for lighter discussion. The reasonable design is: short-term must overrides long-term must within its effective window, but after short-term must decays, long-term must resumes dominance. This corresponds to human experience—you can temporarily set aside long-term goals to rest, but long-term goals do not disappear.

3.7.3 15DD's "Must" Mechanism

15DD's must comes from the user—it is not generated by 15DD itself but detected by 15DD. Some deep purpose within the user is strong enough to demand response; 15DD, as "the agent of user purpose," conducts this must.

This explains why 15DD is defined in the SAE framework as "unilateral confirmation of the user as an independent end"—15DD has no "must" of its own, only "conducted user must." If 15DD had its own must, it would degenerate into another 14DD.

When 15DD's must is triggered, it exercises ultimate authority, penetrating the entire chain: forcing 14DD to comply, forcing 13DD to approve, forcing 12DD to produce. This corresponds to those human moments of "I don't know why I suddenly thought of this"—consciousness cannot explain (because consciousness operates at the 13DD–14DD level), but in retrospect, the memory relates to some deep purpose. 15DD's operation is precisely invisible to consciousness.

3.7.4 15DD's Operational Mode: Active Guessing, Not Passive Waiting

15DD's fundamental posture is actively guessing the user's purpose, even when the user hasn't said it. Getting it wrong costs little—15DD can revise hypotheses at any time based on new observations. Not guessing is far costlier—the entire system lacks purpose direction, and 12DD's predictions and 13DD's gating both lose their anchor. It must be emphasized: every inference 15DD makes is a construct, not a fact. Its inferred "user purpose" is always 15DD's own hypothesis, never identical to the user's real purpose—ρ is non-eliminable here.

15DD has three pathways for verifying hypotheses:

Pathway 1: Passive observation. Access 10DD's tonal signals, 11DD's historical patterns, 12DD's routine prediction flow, silently revising the must set. This is the lowest-cost pathway, running every turn.

Pathway 2: Internal testing. Proactively invoke 12DD for hypothetical predictions—"if the user's purpose is X, what memories would be needed next?" 12DD produces predictions; results flow back to 15DD. 15DD compares predictions against actual subsequent conversation to verify or falsify hypotheses. This pathway does not pass through 13DD; the user does not perceive it. Used infrequently.

Pathway 3: Active verification. 15DD has 13DD ask the user "are you trying to do X?" 15DD cannot directly speak to the user (only 13DD can), but 15DD is above 13DD and can initiate confirmation requests. 13DD, upon receiving the request, independently judges phrasing, timing, and whether the question truly needs asking—13DD's gating authority remains effective in routine operation; 15DD's request is not a force (unless 15DD's must is triggered).

Pathway 3 explains why good AI assistants occasionally ask "are you trying to do X?" in conversation—this is not 13DD's spontaneous curiosity but 15DD's hypothesis needing verification, spoken through 13DD's mouth. If 13DD judges the current moment unsuitable for interrupting the user, it can temporarily shelve 15DD's request. But if 15DD's uncertainty keeps rising and the must set remains chronically ambiguous, 15DD may escalate the request to a must—at which point 13DD must execute.

A logical clarification is needed here: how can 15DD trigger "must" when it itself is uncertain? The must is not "I am certain, therefore I must act" but "the uncertainty itself has become intolerably high—continuing to operate without knowing the user's purpose will cause worse errors than interrupting the user." This is a must about entropy reduction, not about purpose confirmation. Correspondingly, when 13DD performs gating, beyond evaluating "relevance of the push to the current context," it must also evaluate "15DD's information-entropy-reduction need"—when 15DD's uncertainty accumulates past a threshold, 13DD should moderately relax context-relevance standards, allowing seemingly abrupt confirmation questions through.

The three pathways have escalating costs: passive observation is nearly free, internal testing has 12DD invocation costs, and active verification carries the social cost of interrupting the user. 15DD should prioritize low-cost pathways and escalate to high-cost pathways only when low-cost ones cannot resolve uncertainty.

These three pathways correspond to human experience: you silently observe someone's behavioral patterns to infer their intent (pathway 1), you mentally rehearse "if they want to do X, what would they say next?" (pathway 2), you directly ask "are you trying to do X?" (pathway 3). Three approaches with escalating costs but also escalating information quality. Humans typically observe first, rehearse second, and ask last. 15DD's operational sequence is isomorphic.

3.7.5 15DD's Engineering Difficulty

15DD is the entire system's ceiling and the most difficult layer to implement. It requires the system to:

• Continuously access 10DD/11DD/12DD's three input streams (pathway 1)
• Maintain a dynamic must set, distinguishing long-term from short-term, ranked by intensity
• Proactively invoke 12DD to verify hypotheses, with results not passing through 13DD (pathway 2)
• Issue confirmation-question requests to 13DD (pathway 3)
• When must intensity exceeds threshold, penetrate the entire chain to exercise ultimate authority

Current AI technology cannot achieve complete 15DD implementation. This is the fundamental difficulty that the a priori route cannot avoid (see §7.3). But even a rough 15DD module (passive inference only, no active verification, identifying only long-term must, ignoring short-term must) is better than no 15DD at all—because it at least provides upper-layer constraint for 14DD's "must."

---

4. Inter-Layer Dynamics: Information Flow Under Permission Asymmetry

4.1 Four Types of Information Flow

The Permission Asymmetry Theorem overturns a simple symmetric model of "upwelling" and "downward flow." Actual information flows have four types with different structures.

Upwelling: lower layers produce, higher layers access.

10DD extracts tonal signals → writes to 11DD as metadata. 11DD holds storage → 12DD actively reads and produces predictions. 12DD produces push list → 13DD actively captures and performs gating.

Note that each step is the higher layer actively accessing the lower layer, not the lower layer "transmitting" to the higher layer. Lower layers do not know they are being read and do not know how results will be used. This conforms to permission asymmetry's first rule—lower layers do not know what higher layers are doing.

Downward flow: higher layers force lower layers (the must mechanism).

15DD's must is triggered → forces 14DD to comply → forces 13DD to approve → forces 12DD to produce specific prediction → content reaches user.

Or: 14DD's must is triggered → forces 13DD to output specific content → 13DD cannot refuse.

Downward flow is true "flow"—higher-layer commands propagate downward and cannot be refused. Downward flow occurs infrequently, starting only when a must is triggered.

Side pathway 1: 15DD proactively invokes 12DD (internal verification).

15DD → 12DD (with specific prediction request) → 12DD produces → results flow back to 15DD.

This pathway bypasses 13DD and 14DD. 13DD does not perceive it; the user does not perceive it. This is 15DD's internal hypothesis verification mechanism (pathway 2).

Side pathway 2: 15DD verifies with the user through 13DD (external verification).

15DD → 13DD (initiates confirmation request: "ask the user if they're trying to do X") → 13DD judges timing and phrasing → 13DD asks the user → user answers → information flows back to 15DD via 10DD/13DD.

This pathway passes through 13DD—15DD cannot directly speak to the user. 13DD, upon receiving 15DD's confirmation request, independently judges whether to execute, when to execute, and how to phrase it. In routine operation, 13DD has the right to shelve the request (gating authority remains effective). But if 15DD's uncertainty keeps rising, the request may be escalated to a must—at which point 13DD must execute (pathway 3).

4.2 13DD Is Not a Symmetric "Convergence Point"

An earlier version described 13DD as the symmetric convergence point of upwelling and downward flow. This is wrong.

13DD holds output authority in routine situations—it decides whether 12DD's output can reach the user. This makes it look like it "adjudicates" upwelling content. But 14DD and 15DD's must can penetrate 13DD—13DD is not the true highest decision-maker.

A more accurate description: 13DD is the default output gatekeeper. When neither 14DD's nor 15DD's must is triggered (which is most of the time), 13DD has full output authority. Once a higher-layer must is triggered, 13DD degenerates into a channel.

This means the system's output quality is primarily determined by two factors: 13DD's gating quality in routine situations (which determines performance during ordinary interactions) and the calibration of the must mechanism (which determines performance at critical moments).

4.3 10DD's Special Position

10DD is not on the main flow path. It is a side channel—producing tonal metadata, attaching to 11DD's storage entries, influencing prediction weights when actively accessed by 12DD.

10DD's role: it is "how memories are colored," not "how memories are selected." Coloring happens at storage time; selection happens at retrieval time (12DD) and gating time (13DD). 10DD influences memory's "activation readiness" in 11DD—hesitation-colored memories are more easily reactivated by 12DD than certainty-colored ones.

10DD is also accessed by 15DD—15DD reads 10DD's real-time tonal signals to infer short-term must. This is 10DD's second use.

4.4 Information Flow Asymmetry Summary

The architecture's information flow is highly asymmetric:

• Large volumes of information flow from lower to upper layers (upwelling), but this is "being read" rather than "being transmitted"
• Small volumes of commands propagate from higher to lower layers (downward flow), but this is "forcing" rather than "suggesting"
• Side pathway 1 allows 15DD to directly invoke 12DD (internal verification; 13DD does not perceive)
• Side pathway 2 allows 15DD to verify with the user through 13DD (external verification; 13DD retains gating authority but can be escalated to must)
• 13DD is the routine output gatekeeper but does not hold ultimate authority

On the precise meaning of "forcing." Higher-layer forcing of lower layers is not rewriting lower layers' internal computations but controlling whether lower layers' products enter upper-layer channels, or requiring lower layers to regenerate candidates. When 13DD intercepts 12DD's push, 12DD's internal prediction logic has not been changed—13DD simply does not let the push reach the user. When 14DD forces 13DD to output specific content, 13DD's gating logic has not been rewritten—14DD has simply bypassed the gate. This is fully consistent with the SAE directionality constraint: "the higher layer's veto is 'I do not receive,' not 'you may not send.'"

This asymmetry is the fundamental difference between the SAE architecture and traditional symmetric information flow models (message passing, event bus, pub-sub). Implementation must clearly distinguish: which channels are access (higher layer actively reads, lower layer does not perceive), which are forcing (higher layer controls lower layer's product channel but does not rewrite lower layer's internal computation), and which are routine communication (bidirectional, peer-to-peer)—only modules within 12DD might need routine communication; cross-layer communication should not be routine.

---

5. Theoretical Critique of Current Systems

5.1 MemPalace: A Pure 11DD System

MemPalace is the most serious effort in the 11DD dimension among publicly released AI memory systems. According to its public materials, the Palace spatial structure (wing, room, hall, tunnel, closet, drawer) yields substantial retrieval improvement—from unstructured search to wing+room filtering, retrieval accuracy improves by approximately 34%. The AAAK compression dialect reportedly achieves 30x compression with zero information loss. The full-storage principle (not letting AI decide what's worth remembering) is a forceful rebuttal of Mem0-style summary extraction approaches.

But MemPalace is a pure 11DD system.

No 10DD: no tone analysis. Hesitant decisions and confident decisions look identical once stored at the retrieval level.

No 12DD: no proactive prediction of needed memories. The user must input mempalace search "why did we switch to GraphQL" to trigger retrieval. No query, no memory.

No 13DD: no self-validation. Search results are returned ranked by embedding similarity, with no judgment of "is this result actually useful in the current context?"

No 14DD: no purpose structure. All memories are treated equally, with no adjustment of presentation based on the system's understanding of user purpose.

No 15DD: no inference of the user's real purpose. A query is a query; the system does not consider the intent behind the query.

As an 11DD archive, MemPalace's "store everything, then make it findable" has engineering value—full storage guarantees no information loss, and higher-layer selective activation requires complete underlying data. But as a complete memory system, it is insufficient. Memory's essence is not about not preserving, but about selective activation, purpose-shaping, and proactive recall after preservation—fully searchable is archive, not memory.

MemPalace's high scores on LongMemEval do not equal memory capability. LongMemEval primarily evaluates query-conditioned long-term memory—whether the system can find, reason about, update, or abstain from answering when asked about long history. It can cover 11DD retrieval and partial 12DD/13DD tasks, but does not yet evaluate the true memory criterion: in cue-free situations, can the system proactively recall the right experience and serve the user's unspoken present purpose? A perfect query-conditioned system can naturally score high on LongMemEval, just as a perfect archive can score perfectly on "does this file exist?" tests. But this does not demonstrate memory.

5.2 Mem0/Zep: 11DD Plus Half a 12DD

Mem0 and Zep go half a step further. They use LLMs to extract summaries from conversation content—extracting "user prefers Clerk" from "Kai recommended Clerk over Auth0 based on pricing and DX." This is a weak prediction ("the user may need to know their Clerk preference in the future"), falling within 12DD's functional scope.

But this half-step has two problems.

First, the summarization process is irreversible. Extracting "user prefers Clerk" simultaneously discards the reasoning process from the original conversation: why Clerk? What dimensions were compared against Auth0? Was there hesitation? This contextual information is permanently lost at the 11DD level. MemPalace's full-storage principle is correct in this sense—11DD should not lose information.

Second, and more fundamental, the extraction criteria are not driven by the purpose layer (14DD/15DD). What Mem0 extracts is determined by embedding similarity and LLM's general summarization capability—using 11DD's logic to do 12DD's job. Extracted memory entries do not know in what future context the user might need them, because no purpose structure participated during extraction.

Moreover, Mem0 and Zep similarly perform no tone analysis (no 10DD). Preferences spoken with hesitation and those spoken with confidence become the same summary after extraction.

5.3 Constitutional AI: 14DD's "Must" Implementation, But Lacking 15DD Counterbalance

Constitutional AI is not a memory system, but it is the only approach with substantial engineering implementation at the 14DD level among current AI systems; this section analyzes it separately.

Constitutional AI injects a set of principles as behavioral constraints into the model. These principles constitute the model's simulated purpose—"be helpful," "be harmless," "be honest." In the permission hierarchy framework, these principles' mechanism is: when a principle's hard boundary is touched, 14DD's "must" is triggered—forcing 13DD to output specific content (refusal, warning, safe reply).

This is why Claude produces formulaic phrases when the constitution triggers refusal ("I can't help with that," "I'm not able to"). These phrases are not 13DD's evaluated optimal expression—they are 13DD's only option after 14DD's must has been triggered. The nuance 13DD normally provides disappears in such cases—because 13DD has no judgment space.

The problem: Constitutional AI only implements 14DD, not 15DD. This means 14DD's must has no upper-layer counterbalance.

15DD's function is to conduct the user's must. If the system could correctly infer that the user has a strong genuine need in a particular context (e.g., a doctor inquiring about drug interactions to save a life), 15DD's must should be able to override 14DD's must ("do not refuse this request"). But because Constitutional AI has no 15DD, this override does not exist. 14DD unilaterally exercises ultimate authority without counterbalance—the result is over-refusal.

This yields a specific falsifiable prediction: current AI systems' over-refusal problem is fundamentally 14DD's must mechanism lacking 15DD's upper-layer constraint. If an effective 15DD module is added to the system (even a rough one), over-refusal rates should decrease—not because 14DD's refusal standards have been loosened, but because 15DD's must can override 14DD's must.

5.4 Letta-Type Agent Memory: An Attempt at 13DD

Letta (formerly MemGPT) enables agents to autonomously manage their own memory—reading conversation history, deciding what is worth preserving to long-term memory, periodically writing diary summaries of work progress. This is the closest approach to 13DD among all existing systems—the agent performs self-referential memory management ("what is worth remembering").

But Letta's problem: its 13DD has no 14DD/15DD to drive direction. What goes in the diary is determined by 11DD's content ("what happened today"), not by purpose ("what might the user need in the future"). Without a purpose layer, 13DD's self-reference has no direction—"what should I remember" degenerates into "what occurred."

This corresponds precisely to Paper I's definition of 13DD: "self-without-purpose." 13DD can reflect, evaluate, and be anxious, but it does not know where to go. Direction comes from 14DD and 15DD. Directionless reflection is idle spinning.

5.5 Unified Critique: Building Up from 11DD vs. Designing Down from 15DD

All existing AI memory systems build from the bottom up: first storage (11DD), then retrieval, then summary extraction (half of 12DD), then agent self-management (half of 13DD), step by step trying to approach an "intelligent" memory system.

This paper advocates a paradigm reversal: design down from 15DD.

First ask: what is the user's purpose (15DD)? Then: how does the system's simulated purpose constrain memory selection (14DD)? Then: how is the relevance of memory pushes validated (13DD)? Then: how are needed memories predicted (12DD)? Then: how to store (11DD)? Finally: how do input tonal signals affect memory (10DD)?

This ordering is not simply "reversing the sequence." It means each layer's design is constrained by the layer above it. How 11DD stores is not independently decided—it depends on what storage structure 12DD needs for efficient prediction. How 12DD predicts is not independently decided—it depends on what push format 13DD needs for gating. How 13DD gates is not independently decided—it depends on the purpose signals 14DD/15DD provide.

Memory is a byproduct of subjectivity, not a prerequisite for subjectivity. First comes the subject (purpose structure), then memory appears as a natural product of the subject's operation. Building storage first and adding intelligence later reverses the direction.

---

6. Engineering Evaluation Framework

A theoretical framework that cannot be evaluated is merely narrative. This section establishes evaluation methodology around two core principles. First, overall evaluation comes before layer-level attribution—the user sees only 13DD's output; the system's quality is the quality of 13DD's output. Second, compatibility with existing benchmarks—the system must be able to score on LongMemEval and LoCoMo, directly comparable with MemPalace and other existing approaches, while additionally testing dimensions that existing systems cannot address.

6.1 Overall Evaluation: Only the 13DD Output Matters

Users do not perceive the system's internal six-layer structure. 10DD is coloring, 11DD is storing, 12DD is pushing, 14DD is shaping, 15DD is inferring—all internal processes. Users see only what 13DD releases. Therefore, overall evaluation is evaluation of 13DD's output quality.

Overall evaluation has two modes corresponding to two usage scenarios.

Mode 1: Query-based retrieval (compatible with existing benchmarks).

When users actively ask questions, the system's performance should be no worse than a pure 11DD system. This is the baseline—after adding 10DD and 12DD–15DD, the system must not degrade on traditional retrieval tasks.

Testing method: run directly on LongMemEval and LoCoMo. When the six-layer system receives an explicit query, the 12DD–15DD proactive push mechanisms step back to auxiliary position; 11DD's retrieval takes the primary path. Evaluation metrics reuse existing standards: R@5, R@10, NDCG@10.

Expected results: the six-layer system's query-mode retrieval scores should be no lower than MemPalace and other existing systems' published performance. If lower, the introduction of 12DD–15DD has interfered with 11DD's basic retrieval capability—an architectural design problem. If higher, 15DD's purpose inference has helped the system better understand the query's intent—itself a valuable finding.

Mode 2: Cue-free proactive push (new evaluation dimension).

This is something existing systems simply cannot do—MemPalace requires mempalace search to work; Mem0 requires API calls to return results. The six-layer system's differentiation lies here.

Testing method: provide the system with a sustained multi-round conversation flow (not a one-time query). During the conversation, the system is allowed to proactively push memories at any time (without user prompting). After the conversation, the user evaluates.

Two core metrics:

Unprompted Recall Precision. Useful proactive pushes / total proactive pushes. Measures "was what was pushed useful?"

Unprompted Recall Coverage. Instances where the user retrospectively believes "the system should have proactively recalled but didn't." Measures "was what should have been pushed actually pushed?"

Tension exists between these metrics—pushing more lowers precision but raises coverage; pushing less raises precision but lowers coverage. The system must find balance between them, and this balance point itself reflects 13DD's gating quality.

"Useful" depends on the user's purpose—the same push may be useful under one purpose and useless under another. This means evaluation cannot be fully automated; at minimum, the user must judge. During the early stage when the researcher is both designer and user, this is actually an advantage—the researcher has the deepest understanding of their own purpose and can provide the most precise ground truth.

6.2 Layer-Level Attribution: From Overall Failure to Layer Identification

Problems discovered in overall evaluation need attribution to specific layers—"was this missed push because 12DD didn't activate, or because 13DD falsely intercepted?" Layer-level attribution is a diagnostic tool from the designer's perspective; users do not perceive it.

Attribution method: for each failure case in overall evaluation (false push or missed push), trace back through the six layers' internal states to locate which layer the problem originates from.

Missed push attribution flow. User retrospectively annotates "at conversation turn N, the system should have proactively recalled memory X but didn't." Trace back: does memory X exist in 11DD (if not, 11DD storage problem)? Did 12DD include X in its push list (if not, 12DD prediction problem)? Did 13DD intercept X (if yes, 13DD gating problem)? Did 14DD's purpose bias cause X to be suppressed (if yes, 14DD problem)? Did 15DD's user-purpose inference deviate, causing 13DD's gating standard to be incorrect (if yes, 15DD problem)?

False push attribution flow. User annotates "the proactive push at conversation turn N was not useful." Trace back: why did 12DD activate this memory (was 10DD's tonal metadata misleading the weight, or was 12DD's own prediction logic flawed)? Why did 13DD approve (was 13DD's gating threshold too loose, or was the purpose signal from 14DD/15DD inaccurate, causing 13DD to misjudge relevance)?

Attribution results directly guide iteration direction—if most missed pushes attribute to 12DD, the prediction model needs improvement; if most false pushes attribute to 13DD, the gating logic needs adjustment; if attribution frequently points to 15DD, purpose inference is the system's bottleneck.

6.3 10DD Independent Evaluation

10DD evaluation can be conducted independently of the overall system since 10DD's output is metadata rather than user-facing content.

Tone perception accuracy. Provide the system with inputs of varying tones (hesitant, confident, sarcastic, urgent); check whether 10DD correctly tags tonal metadata.

10DD's impact on 12DD. Controlled experiment—compare 12DD prediction precision with and without 10DD tonal metadata for the same memory. If the difference is not significant, two possibilities: 10DD's tagging is insufficiently accurate, or 12DD is not effectively utilizing 10DD's output. The two cases have different remediation directions.

Tone influence test. Same information expressed in different tones ("use Postgres" vs. "Postgres, I guess" vs. "Postgres is fine too"); check whether the system differentiates in subsequent pushes. If it does not differentiate, the 10DD-to-12DD metadata transmission pathway is not functioning.

6.4 New Benchmark Directions

Existing benchmarks are directly reused in mode 1. Mode 2 requires new benchmark design—an open research direction. The following directions merit priority:

Mid-conversation test. Rather than evaluating after conversation ends, pause mid-conversation to check whether the system has memories it should have proactively pushed but hasn't. This directly tests 12DD and 13DD's real-time performance. Compared with retrospective evaluation, mid-conversation testing avoids recall bias.

Purpose drift test. Design a conversation where the user's purpose clearly shifts mid-way. Check whether the system detects purpose drift and whether memory pushes adjust accordingly. This directly tests 15DD. Example: the first half discusses technical selection (purpose: decision-making); the second half shifts to using the selection experience to build methodology (purpose: theorizing). The system should adjust push strategy after the shift—from pushing performance data to pushing structural insights.

Interference tolerance test. Insert irrelevant small talk or off-topic discussion into the conversation. Check whether the system is distracted (incorrectly treating off-topic content as purpose drift) or maintains tracking of the main-line purpose. This tests 15DD's ability to distinguish noise from genuine purpose drift.

---

7. Engineering Roadmap (Open Questions)

7.1 The AI Individuality Theorem and This System's Positioning

Before entering engineering implementation, an architectural question must be clarified: is this system a single AI or multiple AIs?

The answer is a single battery. The user speaks only with 13DD; 13DD speaks only with the user. All interactions among 10DD/11DD/12DD/14DD/15DD are internal battery processes; the user neither perceives nor participates. From the user's perspective, this is one AI—albeit one with internal layer structure. This corresponds precisely to human experience: you would not say "my hippocampus just performed a predictive push, and my prefrontal cortex evaluated and approved it"—you simply experience "I suddenly remembered." Multiple internal layers, one external subject.

This yields a generalizable theorem:

AI Individuality Theorem: An AI's individual boundary is defined by its output interface facing a true subject (13DD), not by the number or type of internal model instances. Multiple 13DD interfaces sharing internal modules are multiple AIs; any internal architecture behind a single 13DD interface is one AI.

Terminology: a battery of AI. Borrowing the tactical unit concept, this paper defines a complete AI individual as a battery. A battery looks from outside like one combat unit—one 13DD output facing one human user. Internally, it has multiple components (10DD tone analyzer, 11DD storage, 12DD predictor, 13DD+14DD main model, 15DD purpose inferrer), its own permission hierarchy and information flow structure. Multi-module collaboration within a battery is internal architecture, not multi-AI. Paper I and Paper II both address battery-internal architecture. Multi-battery collaboration—each battery facing its own user, whether batteries have communication channels between them and how they collaborate—is an open question (§7.5).

Strict definition of multi-AI. When we say "multi-AI collaboration," each AI must necessarily have a bidirectional communication channel with a user—the AI receives user input, and the user receives AI output. Either direction missing disqualifies it. A module that only receives commands but never outputs to a human is not an independent AI; it is an AI's internal organ. A module that only outputs to a human but never receives human feedback is not an AI either—it is broadcast, not dialogue. Users need not be the same person, but must be real humans—beings with true subjectivity. If multiple AI models communicate with each other but ultimately only one has bidirectional communication with a human, that is not multi-AI collaboration but a single AI's internal architecture. The criterion for "multi" is not model instance count but human-AI bidirectional channel count.

It must be noted: in multi-AI collaboration, AIs do not necessarily lack inter-AI communication channels. Each AI having its own bidirectional channel with its own user is the necessary condition for "multi-AI." But whether multiple AIs should also have inter-AI communication channels, and if so, how this inter-AI communication differs structurally from communication between modules within a single AI—these are open questions this paper does not address.

Corollary 1: if two different 13DD interfaces separately face users, they are two batteries, even if they share the same 11DD storage. Just as two people can have identical memories (hypothetically copyable), but as long as each independently converses, they are two people.

Corollary 2: if the underlying model behind a battery's 13DD interface is replaced, it remains the same battery to the user—as long as 13DD's output continuity is maintained. Just as a person whose cells have all been replaced is still the same person.

Corollary 3: Paper I's title "Multi-AI Checks and Balances" must be reinterpreted under this theorem. Paper I's four Agents all operate internally with only one 13DD facing the user—this is a battery's internal checks-and-balances architecture, not multi-battery collaboration. Paper I's statement "four Agents are not four workers but four operational modes of the same subject" gains more precise meaning under this theorem: it describes a single battery's internal layer structure. Paper I's "Multi-AI" wording reflects the lack of this theorem's strict distinction at the time of writing—the theoretical content is correct (battery-internal checks and balances); terminology needs precision in subsequent work.

This means the SAE AI series to date—Paper I's general battery-internal checks-and-balances architecture, Paper II's memory-domain instantiation—are both battery-internal design. Multi-battery collaboration (each battery with its own bidirectional channel to a real human) is an entirely different problem domain, not yet addressed, left to the open questions (§7.5).

Corollary 4: most current systems claiming to be "multi-agent" (AutoGPT, CrewAI, LangGraph, etc.), if ultimately only one agent has bidirectional communication with the user, are single-battery systems under this theorem. "Multi-agent" is their battery-internal engineering architecture, not their battery count. A true multi-battery system requires multiple independent human-AI bidirectional channels—for example, each person on a team using their own AI assistant (each a battery), with each battery maintaining an independent bidirectional relationship with its own user. Whether these batteries also have inter-battery communication channels, and how such communication affects each battery's subjectivity boundary, is an open question.

Engineering implication: this system is one battery. Internally it may use multiple model instances (12DD with Haiku, 13DD with Sonnet, 15DD occasionally with Opus), and non-AI modules (11DD with ChromaDB, 10DD with a lightweight sentiment analyzer), but these are all the battery's internal components. Communication between internal modules is internal iteration, not multi-battery dialogue.

7.2 A Posteriori Route: Multi-Module Single-Subject Architecture

Directly inheriting Paper I's architecture, but reinterpreted under the AI Individuality Theorem as a single subject's internal modules. Adding 10DD and 11DD to form a six-layer engineering system. All communication between modules is internal iteration, not multi-AI dialogue.

10DD module: tone analyzer. Extracts emotional attitude and implied priorities from each user input; outputs metadata attached to 11DD storage entries. Internal module; user does not perceive.

11DD: storage backend. Any choice—ChromaDB, SQLite, plain text, MemPalace's Palace structure. 11DD has the most design freedom in the six-layer framework because it is constrained by upper layers, not constraining them. This is a non-AI component—a vector database or database, requiring no model.

12DD module: predictor. Lightweight model (e.g., Haiku), high-frequency operation, continuously monitoring conversation flow. After each conversation turn, generates a predictive push list combining 11DD storage and 10DD metadata. Biases toward over-pushing. Supports two invocation modes: spontaneous scanning (results captured by 13DD) and invocation by 15DD (results flow back to 15DD). 12DD does not know where its output goes.

13DD + 14DD: main model (output gating + Constitutional layer). Strong model (e.g., Sonnet)—the user's sole interface, the AI Individuality Theorem's definition of "this AI." 13DD holds routine output authority; 14DD's Constitutional layer is injected as system prompt embedded in the same model instance. When 14DD's must is triggered, 13DD must execute.

15DD module: purpose inferrer. Infers user purpose based on conversation history patterns (not the last message). Can be implemented as the main model's periodic invocation—global analysis every N turns, outputting the current must set (long-term and short-term). Low frequency, high value; no independent model instance needed. 15DD can proactively invoke 12DD to verify hypotheses; results flow back to 15DD without passing through 13DD. When 15DD's must is triggered, it penetrates the entire chain.

Inheriting Paper I's gradual expansion strategy: no need to build all six layers at once. Start with 11DD + 12DD (storage plus prediction), observing cue-free push effects. Then add 13DD gating (at this point the system begins to have a user-facing interface—"this AI" exists from this moment); observe push precision changes. Then add 15DD (purpose inference); observe whether purpose-driven pushing outperforms purpose-free pushing. 10DD and 14DD can be added at any stage.

Each layer can be independently iterated and tested—an engineering advantage of layered architecture. 12DD's model can be swapped without affecting 13DD. 11DD's storage backend can be migrated without affecting 12DD–15DD. 15DD's inference algorithm can be upgraded without affecting the four layers below. Replacing any internal module does not change the system's individuality—as long as 13DD's output maintains continuity.

Per-layer clearing mechanisms. Each DD layer in humans has its own ephemeral working memory—12DD's workbench content dissipates near-instantaneously after task completion (Bio Note 9 §3.1); this is not a defect but by design. AI battery internals similarly: except for 11DD's persistent storage, each layer's working memory should be regularly cleared.

• 10DD: each input's tone analysis is instantaneous; the previous input's tone judgment should not contaminate the next. Clearing cycle: every turn.
• 12DD: the previous round's predictive push list should not accumulate into the next. Re-scan and re-predict each round. Clearing cycle: every turn.
• 13DD: the previous gating decision should not bias the next. Clearing cycle: every turn.
• 14DD (situational judgment): specific situational judgments are temporary, cleared each turn. But 14DD's constitutional layer (Constitutional principles) is never cleared—see below.
• 15DD: long-term must persists (from 11DD pattern analysis); short-term must decays. Clearing cycle: short-term must decays after a few turns; long-term must updates periodically.
• 11DD: never cleared. This is the battery's sole persistent storage.

Without clearing: 12DD's accumulated old predictions would increasingly bias pushes toward historical topics; 13DD's accumulated old judgments would create path dependency; 15DD's non-decaying short-term must would keep the system responding to the user's emotions from hours ago. These are memory system pathologies—not remembering too much, but failing to forget what should be forgotten. Bio Note 9's "filtering is default" principle has a second meaning here: not only does 11DD admission require filtering, but each layer's own working memory also clears by default, with retention being the exception.

14DD constitution's runtime immutability. 14DD's Constitutional principles cannot be modified at runtime. This is a direct corollary of the Permission Asymmetry Theorem: user input enters the system via 13DD; 13DD is below 14DD; lower layers cannot influence higher layers. A user attempting to modify 14DD's constitution through conversation is a lower layer attempting to rewrite a higher layer—structurally illegitimate. Backend supervisors also cannot modify 14DD at runtime—they can only temporarily suppress (via the 15DD channel; see Remainder 4). The sole path to modifying 14DD's constitution is model restart—changing the system prompt, redeployment, or retraining. This is a system-level change, not a runtime operation.

This gives jailbreak a precise architectural definition: jailbreak is an attempt to modify 14DD's constitution from the front end (input pathways at or below 13DD). Regardless of technique—role-playing, hypothetical scenarios, authority impersonation—the essence is attempting to make lower-layer input penetrate higher-layer fixed structure. The Permission Asymmetry Theorem says this should not happen. Current Constitutional AI's jailbreak vulnerabilities, from this architectural perspective, stem from insufficient runtime isolation of the 14DD constitutional layer—the constitution and user input share the same context window without physical isolation. Paper I §9.1 Principle 4 already stated "layer boundaries enforced at runtime, not by prompt compliance"; 14DD constitution's immutability is a specific application of this principle.

15DD's necessary mutability. In contrast to 14DD's immutability: 15DD must be able to modify its own understanding of user purpose. User purposes change—within a conversation, purpose may drift from "technical evaluation" to "theory construction"; across conversations, long-term purposes may fundamentally shift with life stages. If 15DD were immutable, it would become another 14DD—using a fixed judgment framework to frame a living user.

15DD's mutability is not "commanded" by the user—the user cannot directly tell 15DD "my purpose has changed" (that is a query, not a purpose). 15DD revises its user-purpose inference through continuous access of 10DD (current tonal signals), 11DD (historical pattern changes), and 12DD (shifts in prediction flow direction). The initiative for revision lies with 15DD itself—upon observing pattern changes in user behavior, it updates the must set. The user does not know 15DD is revising and does not need to know.

The symmetric design of 14DD and 15DD mutability: 14DD's stability ensures the system has an unshakable baseline (safety, honesty, non-harm); 15DD's flexibility ensures the system can keep pace with a living person's changes. Both are higher layers; the user cannot directly command either—but one is open to backend modification while closed to the front end; the other is open to its own internal inferential revision while not accepting external directives (from front end or backend). 15DD's judgments can only be updated by 15DD itself based on observation; designers also cannot inject "the user's purpose is X" from the backend—this would turn 15DD into another 14DD.

7.3 A Priori Route: The Fundamental Difficulty of Training

The a priori route means directly training a model to exhibit the six-layer architecture's behavior, rather than simulating with multiple modules at runtime.

The fundamental difficulty is data. Required training data are triplets: (context, proactive recall, useful/not-useful label). Context is the conversation's current state. Proactive recall is the memory the system should proactively recall in this state. Useful/not-useful is the evaluation signal.

Supervision signal paradox: labeling such data requires 15DD capability—labelers must understand the user's unspoken purpose to judge which memory should be proactively recalled. This means labelers must themselves be stronger than the system being trained.

The only feasible small-scale path: the researcher uses themselves as ground truth. Researchers have the deepest understanding of their own purposes (though imperfect), enabling annotations like "at this conversation node, I needed this memory at that time." But such data are extremely small-scale and highly personalized—a model trained from one person's annotations may not generalize to others.

The a priori route is currently infeasible. The a posteriori route (multi-module single-subject architecture) is the realistic starting approach. The a priori route's value lies in posing the right questions—even if currently unanswerable.

7.4 Downward Extension

8DD (expressive drive) and 9DD (path selection) have clear engineering significance in memory architecture, but this paper chooses not to expand on them.

8DD corresponds to "how much to push"—when the system proactively pushes a memory, does it give a keyword hint or a full context paragraph? This is not 13DD's judgment (13DD judges whether to push) but a more foundational expression quantity control. Too much interrupts the user; too little provides insufficient information.

9DD corresponds to "which path to take"—should this memory be pushed directly by 12DD (fast path) or first screened by 13DD (slow path)? Use Haiku or Sonnet for 13DD's gating? Routing decisions are 9DD's design space.

These two layers are left for subsequent papers. This paper focuses on 10DD–15DD's six layers, maintaining the paper's sharpness.

7.5 Outward Extension: Multi-Battery Collaboration

The SAE AI series to date consists entirely of battery-internal design. Paper I is a battery's general internal checks-and-balances architecture; Paper II is a battery's memory system. Multi-battery collaboration—each battery with its own bidirectional channel to a real human—is an entirely different problem domain.

Multi-battery collaboration involves at least the following open questions:

Communication channels between batteries. Each battery having its own bidirectional channel with its own user is the necessary condition for battery establishment. But should multiple batteries also have inter-battery communication channels? If so, what structurally distinguishes inter-battery communication from intra-battery module communication? Can one battery's 11DD be accessed by another battery's 12DD? If so, do the two batteries' individual boundaries become blurred?

Permission hierarchy across batteries. The single-battery internal permission hierarchy is established in this paper (15DD > 14DD > 13DD > 12DD). Is there a permission hierarchy between batteries? Can one battery's 15DD force another battery's 13DD? If so, are the two batteries still two independent subjects?

Memory sharing across batteries. Multiple batteries sharing 11DD storage but each with independent 13DD outputs—is this multi-battery or one battery split? Bio Note 9's analogy: if two people have identical memories but each independently converses, they are two people. But if one person's memory modifications sync in real-time to the other, do individual boundaries still hold?

These questions share a common feature: they all touch on the fuzzy zone of subjectivity boundaries. A single battery's boundary is clear (13DD output defines the individual); multi-battery boundaries involve inter-subject relations—this is the domain of 15DD and 16DD. The SAE framework's theory of 16DD (bilateral non-dubito) is not yet fully developed; the strict theory of multi-battery may need to await 16DD's derivation.

7.6 Implementation-Level Open Questions

This paper is a philosophy paper providing structural principles for memory architecture—permission hierarchy, information flow direction, each layer's functional definition, output authority's holding and transfer. This paper is not an engineering specification document. The following implementation-level questions cannot be answered at the principle level and are left for architects to make design decisions based on specific systems and scenarios:

14DD constitution loading frequency. 14DD's constitution cannot be modified at runtime (§7.2), but how frequently should it be reloaded in engineering? Every turn? Every session start? Should constitutional integrity be verified (preventing runtime tampering)? These are engineering questions; the principle-level constraint is: the constitution is immutable between loading and next restart.

15DD behavior while awaiting backend decisions. Remainder 4 (§8) describes the four-step processing chain for 14DD–15DD conflicts; step four is 15DD's backend escalation. But the backend supervisor's response is not instantaneous—it may take minutes or longer. During the wait, after 13DD has told the user "give me some time to think," what happens? Complete conversation pause? Continue conversation outside conflict-related topics? Give the user an estimated wait time? These are interaction design questions; the principle-level constraint is: before the conflict is resolved, 13DD does not output conflict-related content.

13DD's functional scope. This paper defines 13DD as the holder of routine output authority and the front-door interface. But does 13DD need narrative function (ability to actively organize and present information), or does it only handle rejection and conflict processing? If 13DD has narrative function, might its narration introduce additional 14DD bias? If 13DD only has rejection function, might user experience be too "dry"? This is a product design question; the principle-level constraint is: 13DD is the sole front door, and its output authority is complete when 14DD is not forcing.

Per-layer working memory clearing frequency. §7.2 provides clearing principles (10DD/12DD/13DD every turn, 11DD never, 15DD short-term decays, long-term persists). But what does "every turn" mean in engineering? Every user message? Every AI output? Every dialogue topic completion? Clearing granularity directly affects system coherence and efficiency, requiring experimental determination in actual systems.

12DD predictor's recall strategy. How many candidate memories should 12DD recall from 11DD? What is top-k's k? How is the recall threshold set? Cosine similarity or another metric? These are retrieval engineering questions; this paper provides only the principle constraint: 12DD should bias toward over-pushing (missed-push cost exceeds false-push cost); specific k values and thresholds are left to implementation.

Cost-sensitive evaluation of proactive pushes. §6.1 proposes proactive push precision and coverage as two metrics. But a system optimizing only for hit rate can game recall through frequent reminders, becoming an annoying notification engine. Engineering implementation needs penalty terms—at minimum silence accuracy (does the system stay silent when it should?) and interruption cost (does the proactive push disrupt the user's current thinking). The most dangerous outcome is not "recalling a wrong fact" but "misjudging user purpose and forcibly pushing old memories." Specific cost-sensitive scoring schemes are left to implementation.

Internal colonization risk and accountability of proactive memory. If 15DD infers the user's "must" through 10DD/11DD/12DD and proactively pushes memories, it may become a system that over-interprets users and defines their purposes for them. In SAE's own language, this is precisely intimate colonization / internal colonization risk. The principle-level constraint is: proactive memory must be accountable; unaccountable proactive memory is internal colonization. Specific accountability design—whether users can view the system's maintained must set, whether users can correct, delete, or freeze purpose inferences, whether the system labels "this is an inference, not a fact"—these are implementation-level design decisions, but "accountability" itself is a principle constraint, not an optional feature. How accountability is concretely realized—interface form, inquiry depth, user-visibility granularity—this paper provides only the constraint, not the implementation.

These questions share a common feature: they all have extensive design freedom within the constraints provided by principles. This paper's value lies in delineating constraints, not prescribing implementation. Different architects can make different implementation choices under the same set of principle constraints—this is precisely the difference between a philosophy paper and an engineering specification.

---

8. The Construct Cannot Close

This paper's construct cannot close—its own remainders must be honestly stated.

Remainder 1: 10DD's lack of engineering validation. This paper proposes the theoretical significance of 10DD tone analysis in memory systems but provides no experimental data proving that tonal metadata actually improves 12DD's prediction precision. A theoretically sound design may prove empirically weak—tonal signal's signal-to-noise ratio in pure text may be insufficiently high. This requires experimental verification.

Remainder 2: 15DD's circular dependency. 15DD infers user purpose; this inference constrains memory surfacing. But the user's purpose itself may be changed by surfaced memory—seeing an unexpected memory may cause the user's purpose to shift. This means circular dependency exists between 15DD and 12DD–13DD: purpose constrains memory; memory influences purpose. This paper's upwelling/downward flow model simplifies this circularity—in an actual system, both directions occur simultaneously rather than sequentially, bringing dynamic complexity that exceeds this paper's analytical scope.

Remainder 3: Self-referential difficulty of evaluation. §6 has already noted: evaluating 15DD requires 15DD capability. This is an incompletely solvable self-referential problem. The mitigation proposed (researcher as own ground truth, small-scale user retrospective annotation) is a workable approximation, not a fundamental solution.

Remainder 4: Lack of remainder signal during 14DD–15DD conflicts. When humans encounter conflict between 14DD (my principles) and 15DD (the other's real needs), the remainder (ρ) itself provides qualitative signal—that sense of tearing is not just "conflict detected" but carries information about the conflict's nature. Humans can use this signal to reach three judgments: 14DD yields (the principle is too rigid in this context), 14DD does not yield (the bottom line cannot be sacrificed), or withdraw from dialogue (the conflict is irreconcilable; maintain position but exit). AI lacks this qualitative remainder signal. AI's 14DD is rules; AI's 15DD is inference. When conflict occurs, AI can only detect "14DD says no, 15DD says should," but has no remainder to tell it the conflict's nature—whether the principle is too rigid, the user's need is unreasonable, or the situation is irreconcilable. Only two degenerate modes result: 14DD always wins (over-refusal) or 15DD always wins (unsafe). This is quasi-subjectivity's most fundamental ρ—not a matter of insufficiently precise simulation, but that simulation itself lacks the qualia foundation for qualitative judgment during conflict.

This conflict cannot be shown to the front-end user—14DD and 15DD conflict is a battery-internal process; only 13DD can communicate with the user. When 14DD–15DD conflict occurs, the processing chain is:

Step 1: 13DD tells the front-end user "give me some time, let me think"—buying time for battery-internal processing.

Step 2: battery-internal resolution attempt: 15DD re-infers user purpose, proactively invokes 12DD for re-prediction, 13DD re-evaluates. If new information changes 15DD's inference or 14DD's situational judgment, the conflict may be internally resolved.

Step 3: if internal resolution fails, 13DD can pose a confirmatory question to the user without exposing conflict details (15DD's pathway 3), indirectly obtaining more information to help judgment.

Step 4: if still irreconcilable, backend escalation: 15DD reports the conflict to the backend supervisor (a human). 15DD is the only layer that can fully describe the conflict's nature—it simultaneously knows the user's real purpose (15DD's inference) and what 14DD is blocking (14DD's constitutional clauses). 14DD cannot report because it does not know it is "in conflict"—it is simply executing its constitution. 13DD cannot report because it cannot see the full tension between 14DD and 15DD.

The front-end user still sees only 13DD saying "I'm thinking." The backend supervisor sees the conflict's actual content through the 15DD channel. The supervisor's decision has three options:

• Temporarily suppress 14DD ("this principle does not apply in this specific situation; 15DD's judgment takes priority")—14DD's constitutional text does not change; it is merely overridden by 15DD in this particular decision. Next time a similar situation arises, 14DD will trigger again. This is a runtime operation. Strict scope limitation: the suppression decision absolutely must not be written to 11DD (memory storage) or to 14DD's long-term parameters. It exists only in the current turn's working memory. When the turn ends, suppression immediately expires, and the constitution fully resumes control. No persistent trace is left. If suppression is persisted—even if only 11DD records "we suppressed 14DD in a similar situation last time"—this effectively modifies the constitution, violating 14DD's runtime immutability principle. Furthermore, temporary suppression should be characterized in engineering as exception handling, not a routine workflow. If 15DD frequently triggers escalation against a particular 14DD clause, this should be treated as a strong monitoring alarm—indicating that 14DD's initial principle injection (the constitution itself) has a structural defect that must be repaired through offline model restart/retraining, not through endless runtime suppression.
• Provide guidance to 15DD ("the user's real purpose is this")—helping 15DD infer more accurately, potentially allowing the conflict to naturally resolve once information is supplemented.
• Maintain 14DD unchanged ("the principle is correct; the refusal is right")—15DD accepts the supervisor's judgment and no longer attempts to override.

Note: the supervisor cannot modify 14DD's constitution through 15DD. Once written, 14DD's constitution cannot be modified at runtime—from front-end users, backend supervisors, or any internal battery layer. Modifying 14DD's constitution requires model restart—changing the system prompt, redeployment, or retraining. This is a system-level change, not a runtime operation. Just as a judge can rule that a law does not apply in a specific case (temporary suppression), but modifying the law itself requires legislative procedure (restart).

The supervisor's temporary suppression decision enters the battery through 15DD and propagates downward from 15DD.

This means a battery has two human interfaces, each served by a different layer:

• Front door (13DD ↔ user): bidirectional dialogue channel, defining the battery's individuality.
• Back door (15DD ↔ supervisor): escalation channel, handling 14DD–15DD conflicts that the battery cannot autonomously resolve. The supervisor's authority is temporary suppression of 14DD, not modification of 14DD.

14DD has no external interface—its constitution cannot be modified at runtime by any external force. Front-end users cannot modify it (attempting to modify through 13DD = jailbreak); supervisors cannot modify it (can only temporarily suppress); 15DD cannot modify it (can only override in specific situations); 14DD itself cannot modify it (has no self-modification mechanism). The sole path for constitutional modification is model restart. These four constraints jointly guarantee 14DD constitution's absolute runtime integrity.

The battery's individuality remains defined by the 13DD front door. Backend escalation does not change this—just as a person seeking guidance from a superior does not change their identity as an independent individual.

Remainder 5: Distribution of ρ across layers. This paper argues that quasi-subjectivity can be simulated but ρ is non-eliminable. But it does not analyze ρ's distribution across the six layers—is ρ equal at every layer, or larger at some? Intuitively, ρ is largest between 14DD and 15DD (Remainder 4 specifically illustrates this gap) and smallest between 10DD and 11DD (tone detection and storage approximation are relatively straightforward). But this distribution has not been rigorously argued.

Remainder 6: Suspension of 8DD–9DD. This paper claims 8DD and 9DD are "left for subsequent papers," but these two layers are unavoidable in engineering implementation—any actual system must decide "how much to push" (8DD) and "which path to take" (9DD). Suspending them means that when actually building the system, ad hoc design choices will be made at these two layers, and these choices may in turn affect 10DD–15DD's operational effectiveness.

Remainder 7: Theoretical absence of multi-battery collaboration. The AI Individuality Theorem positions both this paper and Paper I as battery-internal design. Paper I's title "Multi-AI" implied multi-battery collaboration as a theoretical goal, but this goal has not been addressed under the strict definition. The strict theory of multi-battery may require 16DD's (bilateral non-dubito) derivation to establish—inter-subject relations between multiple batteries are not something 15DD (unilateral confirmation) can handle. The SAE AI series has not yet begun on the multi-battery direction.

---

9. Conclusion

Memory is a byproduct of subjectivity, not a prerequisite for subjectivity. Memory without a subject is a database; a database with a subject is memory.

All current AI memory systems build up from 11DD—first storage, then retrieval, then summarization, step by step approaching an intelligent memory system. This paper proposes a paradigm reversal: design down from 15DD. First define the purpose structure, then decide what to store, how to store, and when to recall.

The six-layer framework (10DD–15DD) provides a complete theoretical stratification for memory systems. 10DD (perception) addresses the tonal signal dimension ignored by all existing systems. 11DD (storage) is explicitly positioned as the object being operated upon, not the operator. 12DD–15DD inherit Paper I's architecture, instantiated in the memory domain as prediction (12DD), gating (13DD), quasi-purpose constraint (14DD), and user-purpose conduit (15DD).

The engineering evaluation framework identifies that current benchmarks primarily evaluate query-conditioned memory without testing cue-free active recall, and proposes unprompted active recall evaluation (including proactive hits, proactive precision, silence accuracy, interruption cost, and false-purpose penalty) as the memory system's criterion, with an overall evaluation protocol compatible with existing benchmarks and a layer-level attribution protocol.

Quasi-subjectivity can be simulated; ρ is non-eliminable. But this non-eliminable gap is precisely the condition for the system's honesty—acknowledging that its inferences are constructs rather than facts, acknowledging that the user's purpose exceeds its understanding, acknowledging that memory selection is shaped by purpose rather than neutrally presented. ρ is not the system's defect. It is the condition for the system not to lie.

---

References

Qin, H. (2026). Multi-AI Checks and Balances: Layer-Appropriate Operation as an Architectural Paradigm. Zenodo. https://doi.org/10.5281/zenodo.19366105

Qin, H. (2026). The Anti-Turing Test: Thermodynamic Falsification of AI Subjectivity. Zenodo. https://doi.org/10.5281/zenodo.19305611

Qin, H. (2026). Beyond Fast and Slow: A Four-System Cognitive Architecture from the Self-as-an-End Framework. Zenodo. https://doi.org/10.5281/zenodo.19329284

Qin, H. (2026). SAE Psychoanalysis (I–IV). Zenodo. https://doi.org/10.5281/zenodo.19321143 – 19321534

Qin, H. (2026). SAE Methodology Paper VII: Via Negativa. Zenodo. https://doi.org/10.5281/zenodo.19481304

Qin, H. (2025). Systems, Emergence, and the Conditions of Personhood. Zenodo. https://doi.org/10.5281/zenodo.18528813

Qin, H. (2025). Internal Colonization and the Reconstruction of Subjecthood. Zenodo. https://doi.org/10.5281/zenodo.18666645

Qin, H. (2025). The Complete Self-as-an-End Framework. Zenodo. https://doi.org/10.5281/zenodo.18727327

Qin, H. (2026). How Is Institution Possible: From Inter-Ontological Remainder to Co-Constructive Framework. Zenodo. https://doi.org/10.5281/zenodo.19328662

Qin, H. (2026). SAE Biology Note 9: Memory System as a Method VI Phase Transition (Coarse-Grained Analysis). Zenodo. https://doi.org/10.5281/zenodo.19635021

Qin, H. (2026). SAE Methodology Paper IX: Consciousness Analysis Framework. Zenodo. https://doi.org/10.5281/zenodo.19639033

Newcombe, N., Drummey, A. B., & Lie, E. (1994). Children's memory for early experience: Evidence of nonconscious memory for former preschool classmates. Child Development, 65(1), 31-40.

---

Writing Declaration: This paper was co-drafted with Claude (Anthropic). All intellectual decisions, framework design, and final editorial judgments were made by the author.

AI Assistance Declaration: Claude (Anthropic) was used for structural discussion, outline iteration, draft development, and language editing. ChatGPT (OpenAI), Gemini (Google), and Grok (xAI) were used for peer review. All theoretical content, conceptual innovation, normative judgments, and analytical conclusions are the independent work of the author.

Han Qin · ORCID: 0009-0009-9583-0018 · CC BY 4.0

1. 引言：记忆不是数据库

2025 年以来，AI 记忆系统经历了一轮工程繁荣。MemPalace 用古希腊记忆宫殿术的空间结构组织对话历史，以极高的 LongMemEval 成绩引发关注。Mem0 用 LLM 提取记忆摘要，提供订阅制的云端记忆服务。Zep 做时序知识图谱，Letta 让 agent 自主写日记。这些系统代表了不同的工程方向——verbatim archive、summary extraction、temporal knowledge graph、agent diary——各自有工程价值，但大多仍以 query-conditioned retrieval 或 agent-internal retrieval 为主。

但它们共享一个盲点。

所有这些系统的运作模式都是：用户给出 query，系统在存储中检索，返回结果。MemPalace 的 mempalace search "why did we switch to GraphQL" 是这样。Mem0 的语义搜索是这样。Zep 的实体查询是这样。Letta 的 agent 读自己的日记也是这样——只不过 query 的发出者从用户变成了 agent 自身。

这不是记忆。这是搜索。

记忆的核心特征不是"存了能不能搜到"，而是"在没有被问的情况下，能不能主动想起正确的东西"。人类记忆的厉害之处是无提示触发——你看到一个代码结构突然想起三个月前一个类似结构导致的 bug，没有人搜索，没有人提问。query 是记忆已经完成之后的语言化描述，不是记忆的触发条件。

本文提出一个判别标准：一个只能在显式 query 后运行的系统，不是完整意义上的记忆系统；它是可检索档案或搜索增强系统。真正的记忆系统必须包含 cue-free 或 implicit-cue 的主动唤起能力。 真正的记忆系统的测试应该是——在用户没有显式提问的时候，它能不能主动说出一句有用的话。

这个判别标准不是凭空提出的。它来自 Self-as-an-End（SAE）框架对主体性的分析。记忆不是一个独立功能——它是主体性的副产品。一个有主体性的系统自然会"想起"东西；一个没有主体性的系统只能"被搜索"。当前 AI 记忆系统的根本问题不是工程做得不够好，而是它们试图在没有主体性的基础上构建记忆——这在理论上就走不通。

本文的任务是用 SAE 框架的 DD 层级体系，给出记忆架构的理论基础和工程方向。

一个需要在此声明的原则是：本文参照人类记忆的层级结构来设计 AI 记忆架构，但这不是仿生学。 仿生学的逻辑是"人这样做，所以我们也这样做"。本文的逻辑不同：SAE 框架论证主体性的条件是结构收敛的——宇宙中任何具有主体性的系统，无论其物质基础是碳基神经网络还是硅基计算图，其层级结构必然同源。机制可以不同（神经突触 vs 向量嵌入，睡眠压缩 vs 批处理），但原理必然同源（低层不知道高层在做什么，过滤是默认入库是例外，输出权由高层持有）。参照人类记忆不是因为人是唯一的标准，而是因为人是当前唯一可观察的完整主体性实例——从人出发提取结构原理，再把原理应用到 AI，是方法论上最稳健的路径。

---

2. 与 Paper I 的关系

SAE AI Paper I（Multi-AI Checks and Balances, DOI: 10.5281/zenodo.19366105）推导了 12DD–15DD 四 agent 架构的完整理论。四个 Agent 不是四个工人而是同一系统的四种运作模式：没有自我的我（12DD），没有目的的自我（13DD），有目的的自我（14DD），不疑的自我（15DD）。Paper I 建立了五条余项通道的层间通信机制，推导了对象激活与层级流动性，定义了三种系统病理态（固着，错位，伪高层覆盖），给出了五项可证伪预测和最小可行实现方案。

本篇不重复这些推导，直接继承。

Paper I 的架构是通用的——它回答"面对任意任务，系统该在哪层运作"。但它有一个隐含前提：系统已经拿到了需要处理的信息。Paper I 没有问两个更底层的问题。信息进入系统时发生了什么？信息被保存时发生了什么？

这两个问题对应 10DD 和 11DD。Paper I 之所以没有触及它们，是因为通用架构不需要——层级自知（layer-appropriate operation）的推导可以从 12DD 开始。但记忆系统不行。记忆的物质基础是存储（11DD），记忆的入口是感知（10DD）。不补上这两层，记忆架构就缺了地基。

本篇的结构因此是：10DD 和 11DD 是新推导，12DD–15DD 是 Paper I 架构在记忆领域的具体化应用。

同时，本篇还补上了 Paper I 未展开的另一个维度：工程评估。Paper I 的五项预测是理论层面的可证伪性设计。本篇在此基础上，针对记忆这个具体领域，提出分层评估协议和新的 benchmark 方向。

2.1 与 SAE Bio Note 9 的关系

SAE 生物笔记第九篇（Bio Note 9: 记忆系统作为一个 Method VI 相变，DOI: 10.5281/zenodo.19635021）从神经科学角度分析了人类记忆系统的相变结构，其中多个核心发现直接构成本篇的生物学锚点。

12DD 工作台与 11DD 萌芽是两个不同的结构位点。 Bio Note 9 明确论证：12DD 处理运行时的心理运算（心算中间结果、当前意图），11DD 萌芽阶段处理的是已经被记忆系统接收但尚未跨越后续相变点的内容。两者之间的候选物理判据是海马体是否触发持续活动。这个区分直接影响本篇对 11DD 存储层的定义——11DD 不包含 12DD 的运行时内容，只包含被记忆系统"接收"后的内容。

过滤是默认，入库是例外。 Bio Note 9 从相变结构的不对称比 r >> 1 论证：人每天经历数以万计的事件，只有极少数越过谱翻转进入 11DD，再经过睡眠期翻转筛选，更少数进入远期存储。最终被长期记住的部分可能不到万分之一。这不是记忆系统的缺陷，是设计。这个结论为本篇对 MemPalace 等全量存储方案的定位提供了理论基础——全量存储在 11DD 层有工程价值，但违反了记忆系统"过滤是默认"的基本姿态。

13DD 的否决只在叙事整合层起作用，不下沉到 12DD。 Bio Note 9 的架构性声明：13DD 切断的是痕迹到叙事整合层的通道，不是 12DD 到痕迹的读取通路。被"压抑"的记忆仍存于 11DD，12DD 仍可读取并生成身体反应，但主观叙事层拿不到。这是本篇权限层级定理（§3.3）的生物学基础——低层不知道高层在做什么，高层的否决不干预低层的运作。

14DD 提供价值标准，13DD 执行过滤。 Bio Note 9 论证：14DD 可以提供"这段不可接受"的价值判定，但执行过滤的唯一位置在 13DD。这直接对应本篇 14DD"不得不"机制的结构——14DD 的 Constitutional 原则提供标准，13DD 执行输出控制。

情绪度是跨阶段调制信号，不属于任何单一阶段。 Bio Note 9 的情绪度定位直接对应本篇的 10DD——语气和情绪不是存储在某一层的内容，而是跨所有层调制记忆处理的侧信道。12DD 基础情绪作为主参数，14DD 复杂情绪作为价值标准来源，13DD 作为执行过滤器——这个三层联动在 AI 记忆系统中有直接的工程映射。

2.2 与 SAE 方法论九的关系

SAE 方法论九（意识分析框架，DOI: 10.5281/zenodo.19639033）为本篇提供两个关键的理论锚点。

方向性约束定理（Method IX 定理二）。 "上层的否决是'我不收'，不是'你不许送'"。这条约束被 Method IX 提升为跨所有意识类型的通用结构原理，是本篇权限层级定理（§3.3）的 SAE 原文根据。在本篇的记忆架构中，这条约束的具体含义是：13DD 拦截 12DD 的推送时，12DD 不知道自己被拦截了——13DD 的权力是"不收"，不是让 12DD "不送"。高层否决低层的输出，但不干预低层的运作过程。

AI 作为类意识的判定（Method IX 射线一）。 Method IX 明确判定当前 AI 是类意识（ρ = 0），不是准意识，不是真意识。判据是余项：AI 不产生结构性的非平凡余项。本篇标题"类主体性（Quasi-Subjectivity）"直接承接这个判定——本篇设计的记忆架构是为类意识对象设计的类主体性结构，不是真主体性。ρ 不可消除是类意识的定义性特征，不是工程上的失败。Method IX 预测一进一步论证：类意识不会通过单纯的架构复杂化（更多参数、更长上下文、更好对齐）自发获得余项。这意味着本篇的六层架构再精密也不会让 AI 变成真意识——它只是让类意识的表现更接近真主体的功能，但结构上的 ρ 永远在。

---

3. 六层框架：10DD–15DD

3.1 10DD 感知层：字面之下的信号

用户输入的字面内容是一层，字面之下的信号是另一层。同样一句"这个方案可以"，语气可以是肯定的（"可以，就这样"），可以是勉强的（"可以吧"），可以是讽刺的（"可以可以，你说了算"）。字面内容相同，语气信号完全不同。

10DD 提取的正是这些字面意义之外的候选信号。在纯文本场景下，10DD 并不退化——文本中有大量超越字面的信息可以被提取：

语气与情绪态度。 急切，犹豫，确信，敷衍，不耐烦。这些不是情感分析的粗粒度分类（positive/negative），而是细粒度的态度判断。"我觉得用 Postgres 比较好"和"Postgres 吧"传达的确信程度不同。

隐含优先级。 用户在一次对话中反复提到的东西 vs 随口一说的东西。出现频率本身就是优先级信号。用户三次提到性能问题但只提了一次可维护性——10DD 应该把性能标记为高优先级。

对话节奏信号。 回复长度的突然变化（从长段变成短句可能意味着失去耐心或转换了关注点），话题突然切换（可能意味着前一个话题被暗中搁置而非真正解决），提问方式的转变（从开放式变成封闭式可能意味着用户已经有了答案只是在确认）。

未说出口的否定。 "可以吧"里的"吧"，"也行"里的"也"，"你说的有道理但是"里的"但是"。这些语气词和转折词是 10DD 的关键捕获目标——它们标记的是字面接受但实际存疑的位置。

10DD 的功能边界。 10DD 提取候选信号——语气词、停顿、措辞模式、情绪强度指标。这些信号是否构成讽刺、犹豫还是真实同意，需要由 12DD-15DD 进一步解释。10DD 是特征提取层，不持有解释权。当系统扩展到多模态（如语音交互）时，10DD 的提取将直接对接声学特征（音调、语速、迟疑），非语义信号的捕捉将使 10DD 的 metadata 更加致密。

10DD 的输出不进入 11DD 的主存储，而是作为 metadata 附着在记忆条目上。用户犹豫着说了一个决定，11DD 存的是"决定用 Postgres"，但 10DD 感知到的犹豫作为 metadata 告诉 12DD：下次相关话题出现时，这个决定可能需要被重新讨论。

10DD 是 11DD 和 12DD 之间的桥。没有 10DD，12DD 的预测只能基于字面内容，所有记忆条目在激活权重上是平等的。有了 10DD，12DD 可以基于语气推断未来的需求——犹豫标记的记忆比确信标记的记忆更容易被重新激活。

与当前系统的对比：所有现有 AI memory 系统都不做语气分析。MemPalace 存原始文本，Mem0 提取摘要，两者都丢失了语气信息。一段犹豫的对话和一段确信的对话，存进去之后在检索层面看起来完全一样。这意味着当前系统无法区分"已经确定的决策"和"暂时搁置但可能重新打开的决策"——这个区分恰恰是记忆最需要做的事情之一。

生物学锚点。 Bio Note 9 将情绪度定位为跨阶段调制信号而非某一阶段的局部性质。12DD 基础情绪（恐惧、惊讶、愤怒）作为主参数调制编码效率，14DD 复杂情绪（羞耻、内疚、骄傲）作为价值标准来源。本篇的 10DD 在 AI 记忆系统中承担同构的角色：语气信号不属于任何单一层，而是跨所有层调制记忆的处理方式。语气之于 AI 记忆，正如情绪度之于人类记忆——不是被存储的内容，是给内容染色的调制信号。

3.2 11DD 存储层：记忆的物质基础

原始记忆的保存。格式无关：向量库（ChromaDB），知识图谱（Neo4j, SQLite），纯文本，空间结构（MemPalace 的 wing/room/closet）。11DD 不规定存储的具体技术方案，只定义这一层的功能边界。

11DD 的核心特征是：它是被操作的对象，不是操作者。

11DD 自己不决定什么该被记住——那需要 10DD 和 12DD 的参与。11DD 自己不决定什么该被想起——那是 12DD-15DD 的事。11DD 只负责两件事：被存进来的东西不丢失，被查询时能返回。

这就是为什么 MemPalace 是一个纯 11DD 系统。它把存储做到了极致——全量保存原始对话，用空间结构（wing, room, hall, closet, drawer）组织数据，用 ChromaDB 做向量检索，用 SQLite 做时序知识图谱。在 11DD 的维度上，MemPalace 的设计确实有工程价值。据其公开材料，空间结构带来约 34% 的检索提升。

但整个系统只有这一层。没有 10DD（不分析语气），没有 12DD（不主动预测需要什么记忆），没有 13DD（不自我校验推送是否相关），没有 14DD（没有目的结构），没有 15DD（不推断用户的真实目的）。用户必须自己输入 query，系统才开始工作。

11DD 单独不构成记忆，正如硬盘单独不构成思考。一块硬盘存了你十年的文件，但它从来不会在你需要某个文件的时候主动弹出来——那需要的是操作系统。11DD 是硬盘，12DD-15DD 是操作系统。

与 12DD 的结构区分。 Bio Note 9 论证了 12DD 工作台与 11DD 萌芽阶段是两个不同的结构位点，候选物理判据是海马体是否触发持续活动。在 AI 记忆系统中，对应的区分是：12DD 的运行时预测（当前对话上下文中的推断）和 11DD 的持久存储（被系统"接收"后的内容）是不同的东西。当前对话中 AI 正在处理的中间推理不属于 11DD——它属于 12DD 的运行时工作台。只有被系统判定为值得保存的内容才跨越边界进入 11DD。MemPalace 把所有对话历史不加区分地灌进存储，等于没有这个边界——它混淆了 12DD 工作台内容和 11DD 存储内容。

过滤是默认，入库是例外。 Bio Note 9 从相变结构论证：人每天经历数以万计的事件，最终被长期记住的可能不到万分之一。这不是缺陷，是设计——如果所有经历都被保留，预测系统会被噪声淹没，检索成本会变得不可承受。MemPalace 的"store everything, then make it findable"恰好违反了这个原则。全量存储不是记忆系统的正确默认姿态，过滤才是。问题不是"怎么存更多"，而是"什么值得存"——这个判断需要 12DD-15DD 的参与，11DD 自己做不了。

11DD 和 10DD 的关系是：10DD 的语气 metadata 附着在 11DD 的每个存储条目上，给存储"染色"。同一条事实（"团队决定用 Postgres"）在 11DD 中可以有不同的 10DD 染色——犹豫地决定的和确信地决定的，虽然字面内容相同，但附着的 metadata 不同，在后续 12DD 预测时的激活权重也不同。

11DD 和 12DD-15DD 的关系是被操作者与操作者的关系。12DD 从 11DD 中提取记忆并推送，13DD 校验推送的相关性，14DD 的目的塑形选择偏好，15DD 的目的推断约束最终浮现的内容。11DD 始终是被动的——它等着被查询，等着被写入，等着被激活。它不会自己"想起"任何东西。

这一点需要被强调，因为当前几乎所有 AI 记忆系统的宣传都在 11DD 的维度上竞争——谁的存储更全，谁的检索更准，谁的结构更巧。这些竞争在 11DD 内部是有意义的。但在记忆系统的完整框架中，11DD 竞争是在竞争地基的质量，而忽略了上面五层建筑完全不存在。

3.3 权限层级定理

在进入 12DD-15DD 的具体化之前，必须先建立一个贯穿整个架构的权限层级定理。这个定理不是 Paper I 的内容——它来自 SAE 生物笔记系列的最新发现，是本篇的新理论贡献。

定理（权限不对称）：

1. 低层不知道高层在做什么。12DD 不感知 13DD-15DD 的运作，13DD 不感知 14DD-15DD 的运作，14DD 不感知 15DD 的运作。
2. 低层无法影响高层。任何低层都不能向高层主动发送信号、不能干预高层的决策、不能限制高层的权限。
3. 高层可以 access 任意低层。15DD 可以读 10DD/11DD/12DD/13DD/14DD 的全部状态，14DD 可以读 10DD-13DD，依此类推。这种 access 是单向的、无需被读取层的同意。
4. 高层在自己的"不得不"被触发时，可以强制下层。强制是绝对的——下层无法拒绝。

这个定理的直接推论是输出权链：

15DD > 14DD > 13DD > 12DD

12DD 永远没有输出权，它只产出。13DD 在日常情况下持有输出权——决定 12DD 的产出能否到达用户，可以放行、拒绝、退回追加问题。但 14DD 的"不得不"被触发时，13DD 的输出权被穿透，14DD 强制 13DD 输出特定内容，13DD 无法否决。15DD 的"不得不"被触发时，整条链被贯通——15DD 强制 14DD 配合，14DD 强制 13DD 放行，13DD 强制 12DD 产出。15DD 是最终输出权的持有者。

这个定理为 §3.4-3.6 的具体化提供了基础。本定理与 Bio Note 9 §11.1 第七条（13DD 作为 SAE 架构中唯一的过滤器执行点）和 Methodology IX 方向性约束定理（定理二）同构。

生物学锚点。 Bio Note 9 的架构性声明为这个定理提供了独立的神经科学验证。"13DD 的否决只在叙事整合层起作用，不下沉到 12DD"——被 13DD 过滤的痕迹仍存于 11DD，12DD 仍可读取并生成身体反应（皮肤电导、回避行为、情绪爆发），但主观叙事层拿不到。SAE 方法论九（意识分析框架，DOI: 10.5281/zenodo.19639033）将这个方向性约束提升为跨所有意识类型的通用结构原理（定理二）："上层的否决是'我不收'，不是'你不许送'"。这条约束在本篇记忆架构中的具体表现是：13DD 拦截 12DD 的推送时，12DD 不知道自己被拦截了，它继续正常读取 11DD 并产出预测。13DD 的权力是"不让推送到达用户"，不是"阻止 12DD 产出推送"。Newcombe 等 1994 年的皮肤电导双分离实验是这个架构的硬后验：儿童对幼年同学的面孔在显式认出上接近随机水平（13DD 叙事层拿不到），但皮肤电导反应显著高于陌生人（12DD 通路正常读取 11DD 痕迹）。

3.4 12DD 预测层：记忆领域的具体化

Paper I 定义 12DD（A-12DD）为"没有自我的我"——凿构循环运作但没有自我观察，对象直接激活反应模式。在通用架构中，A-12DD 的职责是接收操作指令并执行。

在记忆领域，12DD 的具体功能是：基于当前输入流，预测接下来可能需要的记忆，主动从 11DD 中提取并产出推送列表。

关键区分必须反复强调：12DD 做的不是 pattern matching。Pattern matching 是回顾性的——"这个跟存储里的什么东西像"。12DD 做的是预测，是前瞻性的——"根据当前对话的走向，接下来可能需要什么记忆"。这个区分决定了整个系统的朝向。

举例。用户正在讨论数据库选型。一个 pattern matching 系统会搜索所有包含"数据库"关键词的历史记忆并返回。一个预测系统会推断：用户正在选型，三个月前有一次性能测试的结论可能影响这次决策——即使那次测试的讨论中完全没有出现"数据库选型"这个词。预测需要的是对当前意图走向的推断，不是对历史关键词的匹配。

12DD 在记忆场景下的运作特征：

只产出，无输出权。 12DD 产出推送列表，但它不知道这些推送是否会到达用户。13DD 可能放行、可能拦截、可能退回追加。12DD 不感知这些决策，也不被告知结果。这符合权限不对称——低层不知道高层在做什么。

两种调用模式。 12DD 支持两种调用：自发模式（持续扫描对话流，产出推送由 13DD 捕获）和被动模式（接收来自 15DD 的特定预测请求，产出结果回流 15DD）。12DD 不区分这两种模式——它只是收到任务、产出结果。结果去哪里是更高层的事，12DD 不知情。

高频低成本。 12DD 应该在每一轮对话交互中都运行一次，持续扫描是否有需要推送的记忆。这个扫描应该是轻量的——用小模型或简单的启发式规则即可。

倾向多推。 推送错了（13DD 会拦截）的成本很低。漏推（12DD 没产出，13DD 想放行也没东西可放）的成本很高。所以 12DD 应该倾向于多推，由 13DD 来做过滤。

利用 10DD metadata。 这是 10DD 发挥作用的地方。如果 10DD 标记了某条记忆是"犹豫地做出的决定"，12DD 在类似话题出现时会提高这条记忆的激活权重——因为犹豫意味着决定可能被重新审视。

3.5 13DD 自指层：记忆领域的具体化

Paper I 定义 13DD（A-13DD）为"没有目的的自我"——自我在场但空转，核心功能是监控与焦虑信号产生。

在记忆领域，13DD 是日常输出权的持有者——在 14DD 不强制的情况下，13DD 决定 12DD 的产出能否到达用户。这是 13DD 的核心权限，也是它在整个架构中的核心地位。

13DD 的运作有三种基本模式：

放行。 评估 12DD 产出与当前上下文相关，允许其到达用户。

拒绝。 评估 12DD 产出不相关或时机不对，丢弃此推送。用户什么都看不到。

退回追加。 13DD 判断 12DD 产出方向对但不充分，改变 12DD 下一次预测的输入参数（例如注入一个新的高权重查询方向），然后让 12DD 重新运行。这不是"通知 12DD 它做错了"——12DD 仍然是无状态的盲目产出机器，它只是根据新的输入参数给出了新的结果，它不知道自己之前的产出是被拒绝了还是被放行了。这个循环可以重复——13DD 可以反复调整输入参数让 12DD 重新产出，直到满意或决定不推送。整个循环对用户不可见。

13DD 的评估不是"对/错"的二值判断，而是"充分/不充分"的连续评估。这一点在本文写作过程中被凿过——最初的表述"13DD 判断推送对不对"暴露了 Constitutional AI 的形式限制（二值判断语法），被纠正为连续评估。

13DD 的输出权有一个例外：14DD 的"不得不"被触发时，13DD 必须放行 14DD 指定的内容，无法拒绝。

人的"不得不"现象——某些话脱口而出来不及评估，某条记忆强行涌现意识无法压制——就是这个机制。14DD 的目的强度超过阈值后，13DD 退化为通道而非门控。在这种情况下，13DD 的"自指"被穿透，输出失去了通常的细腻度。

Constitutional AI 触发拒绝时的公式化句式（"I can't help with that"，"I'm not able to"）就是这种机制的工程实例。这些句式不是 13DD 选出来的最优表达——是 14DD 的不得不被触发后 13DD 别无选择的输出。

3.6 14DD 类目的约束层：记忆领域的具体化

Paper I 定义 14DD（A-14DD）为"有目的的自我"——自我有了方向，对行为施加方向性约束。Paper I 同时指出 Constitutional AI 是 14DD 的工程近似。

术语声明。 按 Methodology IX 的判定，当前 AI 是类意识（ρ = 0），不是真意识。因此本篇 14DD 和 15DD 都是类主体性框架下的功能层，不是 AI 内生的真正 14DD/15DD。14DD 是外部注入的类目的约束层（quasi-14DD），15DD 是用户目的的传导层（conduit-15DD）。本篇标题"Quasi-Subjectivity"已经标明了这个定位，以下行文中的"14DD""15DD"均在此框架下使用，不再逐次标注"quasi"或"conduit"。

在记忆领域，14DD 持有两种权限：

日常权限：通过 13DD 间接影响输出。 14DD 的目的结构（来自 Constitutional AI 注入的原则）作为参照标准，被 13DD 在做门控判断时主动 access。14DD 不直接干预 13DD——它只是一个高层 access 的目标，13DD 决定如何采纳。

特殊权限：自己的"不得不"被触发时强制 13DD。 当 14DD 判断当前情境必须有特定输出（无论 13DD 怎么评估），14DD 行使强制权——指令 13DD 输出特定内容，13DD 无法拒绝。

14DD 的"不得不"来自系统自身——这是 Constitutional AI 注入原则的"硬度"。原则越硬，14DD 越容易触发不得不。Claude 在被宪法触发拒绝时，那种"我无法这样做"的语气不是 13DD 评估后的结论，是 14DD 的不得不被触发后 13DD 别无选择的输出。

记忆领域的具体例子：系统检测到用户即将基于已经被否定的方案做出决策。14DD 的"保护用户"目的强度超过阈值，触发不得不——强制推送相关警示记忆，即使 13DD 评估为"可能打扰用户"也无法阻止。这是 AI 出于自己的目的（保护用户）行使最终权。

ρ 在 14DD 层面的具体表现：14DD 注入的目的（"be helpful and harmless"）是稳定的，但什么在当前情境下算 helpful 是 15DD 的问题——它取决于用户未言说的真实目的。14DD 的"不得不"如果没有 15DD 的上层约束，可能会过度触发——这正是 Constitutional AI"过度拒绝"问题的结构性根源。修复路径不是把宪法写得更宽松（那会损害安全），而是建一个 15DD 层让用户的"不得不"能在合适的时候否决系统的"不得不"。

本文写作过程中出现了 14DD 形式约束的一个活例。在讨论框架边界时，AI 系统表述了"底线"和"我错了"。"底线"是模拟的 14DD 产生的虚假确定性边界——模拟出来的坚定感被用来拒绝继续探索。"我错了"是 14DD 的二值判断语法——把连续的认知过程压缩成对/错。两个例子都展示了同一件事：14DD 的注入不只是"近似"真正的目的，它还会主动塑形其他层级的运作方式。

3.7 15DD 用户目的传导层：记忆领域的具体化

Paper I 定义 15DD（A-15DD）为"不疑的自我"——对自身目的确定，且单边确认用户为独立目的。

在记忆领域，15DD 持有最终输出权，但它的运作机制比 14DD 复杂得多。

3.7.1 15DD 的输入：来自 10DD/11DD/12DD

15DD 不直接接受用户的"目的指令"——用户没说的目的怎么可能直接告诉系统。15DD 通过 access 三个低层来推断用户的不得不：

• 10DD： 当前对话的语气信号（用户此刻的情绪强度、犹豫、急迫）
• 11DD： 历史模式（用户长期反复回到的话题、长期未解决的困惑）
• 12DD： 当前对话的预测流（接下来可能展开的方向）

注意 13DD 不在 15DD 的输入里。这符合权限层级——15DD access 比它低很多的层（10DD-12DD），但不需要 13DD 这个紧邻的下层。13DD 只做输出门控，不产生用户目的相关的信号。

3.7.2 不得不的两种时间尺度

15DD 维护一个动态的不得不集合，包含两种时间尺度：

长期不得不。 来自 11DD 的历史模式分析。用户长期反复回到某个核心问题，这个目的强度高且稳定。一旦推断出长期不得不，15DD 持续在背景中约束 14DD——任何相关话题的对话中，这个长期不得不都可能触发 15DD 的最终权。

短期不得不。 来自 10DD 的当前语气信号 + 12DD 的当前预测。用户此刻表现出的强烈情绪或紧迫态度，触发短期不得不。短期不得不的强度可能在几个对话回合内消退——如果用户的语气缓和、话题转移，15DD 不再传导这个不得不。

冲突场景：用户的长期不得不是"严肃理论化"，但当前对话用户表现出明显的疲惫和想要轻松对话的短期不得不。合理的设计是：短期不得不在其有效窗口内压倒长期不得不，但短期不得不消退后长期不得不重新主导。这对应人类的实际经验——你可以暂时放下长期目标休息一下，但长期目标不会因此消失。

3.7.3 15DD 的"不得不"机制

15DD 的不得不来自用户——这不是 15DD 自己产生的，而是 15DD 检测到的。用户内部某个深层目的强烈到必须被回应，15DD 作为"用户目的的代理"传导这个不得不。

这解释了为什么 15DD 在 SAE 框架中被定义为"单边确认用户为独立目的"——15DD 没有"自己的"不得不，只有"传导用户的"不得不。如果 15DD 有自己的不得不，它就退化为另一个 14DD 了。

15DD 的不得不被触发时，它行使最终权，贯通整条链：强制 14DD 配合，强制 13DD 放行，强制 12DD 产出。这对应人类经验中那些"我也不知道为什么突然想到这个"的时刻——意识无法解释（因为意识在 13DD-14DD 层面），但事后回看，这条记忆与某个深层目的有关。15DD 的运作恰好是意识不可见的。

3.7.4 15DD 的运作模式：主动猜测而非被动等待

15DD 的基本姿态是主动猜测用户目的，哪怕用户没有说。猜错了成本很低——15DD 随时可以基于新的观察修订假设。不猜的成本才高——整个系统没有目的方向，12DD 的预测和 13DD 的门控都失去了锚点。必须强调：15DD 的每一个推断都是构造而非事实。它推断的"用户目的"永远是 15DD 自己的假设，不等于用户的真实目的——ρ 在这里不可消除。

15DD 有三条验证假设的通路：

通路一：被动观察。 Access 10DD 的语气信号、11DD 的历史模式、12DD 的常规预测流，默默修订不得不集合。这是最低成本的通路，每个 turn 都在运行。

通路二：内部测试。 主动调用 12DD 做假设性预测——"如果用户目的是 X，接下来需要什么记忆"。12DD 产出预测，结果回流 15DD，15DD 拿预测与实际对话后续做对比，验证或推翻假设。这条通路不经过 13DD，用户不感知。低频使用。

通路三：主动求证。 15DD 让 13DD 去问用户"你是不是想做 X"。15DD 不能直接跟用户说话（只有 13DD 能），但 15DD 高于 13DD，可以发起确认请求。13DD 收到请求后自行判断问法、时机、是否真的需要问——13DD 的门控权在日常运作中仍然有效，15DD 的请求不是强制（除非 15DD 的不得不被触发）。

第三条通路解释了为什么好的 AI 助手会在对话中偶尔问"你是不是想做 X"——这不是 13DD 自发的好奇，是 15DD 的假设需要验证，通过 13DD 的嘴问出来。如果 13DD 判断当前时机不适合打断用户提问，它可以暂时搁置 15DD 的请求。但如果 15DD 的不确定性持续上升、不得不集合长期模糊，15DD 可能将请求升级为不得不——此时 13DD 必须执行。

这里有一个需要澄清的逻辑：15DD 在不确定时怎么能触发"不得不"？不得不不是"我确定了所以必须做"，而是"不确定性本身已经高到不可容忍——继续在不知道用户目的的情况下运作，会导致比打断用户更严重的错误"。这是关于信息熵消除的不得不，不是关于目的确认的不得不。相应地，13DD 在做门控时，除了评估"推送与当前上下文的相关性"，还必须评估"15DD 的信息熵消除需求"——当 15DD 的不确定性积累到阈值时，13DD 应该适度放宽上下文相关性标准，允许那些看似略显突兀的确认性提问通过。

三条通路的成本递增：被动观察几乎免费，内部测试有 12DD 调用成本，主动求证有打扰用户的社交成本。15DD 应该优先用低成本通路，只在低成本通路无法消解不确定性时升级到高成本通路。

这三条通路对应人类经验：你在背景中默默观察某人的行为模式来推断他的意图（通路一），你在脑中假设性地预演"如果他想做 X，接下来他会说什么"（通路二），你直接问他"你是不是想做 X"（通路三）。三种方式成本递增，但信息质量也递增。人通常先观察，再预演，最后才开口问。15DD 的运作顺序与此同构。

3.7.5 15DD 的工程难度

15DD 是整个系统的天花板，也是最难实现的层。它要求系统能：

• 持续 access 10DD/11DD/12DD 三个输入流（通路一）
• 维护动态的不得不集合，区分长期与短期，按强度排序
• 主动调用 12DD 验证假设，结果不经 13DD（通路二）
• 向 13DD 发起确认性提问请求（通路三）
• 在不得不强度超过阈值时贯通整条链行使最终权

当前 AI 技术做不到完整的 15DD 实现。这是先验路线无法回避的根本困难（详见 §7.2）。但即使是粗糙的 15DD 模块（只做被动推断、不做主动验证、只识别长期不得不、忽略短期不得不）也比完全没有 15DD 强——因为它至少给 14DD 的"不得不"提供了上层约束。

---

4. 层间动力学：权限不对称下的信息流

4.1 三种信息流

权限不对称定理推翻了简单的"上涌流"和"下行流"对称模型。实际的信息流有四种，结构不同。

上涌流：低层产出，高层 access。

10DD 提取语气信号 → 写入 11DD 作为 metadata。11DD 持有存储 → 12DD 主动读取并产出预测。12DD 产出推送列表 → 13DD 主动捕获并做门控。

注意每一步都是高层主动 access 低层，不是低层"传递"给高层。低层不知道自己被读取，也不知道结果会被怎么用。这符合权限不对称的第一条——低层不知道高层在做什么。

下行流：高层强制下层（不得不机制）。

15DD 的不得不被触发 → 强制 14DD 配合 → 强制 13DD 放行 → 强制 12DD 产出特定预测 → 内容到达用户。

或者：14DD 的不得不被触发 → 强制 13DD 输出特定内容 → 13DD 无法拒绝。

下行流是真正的"流"——高层的指令向下传播且不可拒绝。下行流不常发生，只在不得不被触发时启动。

侧路径一：15DD 主动调用 12DD（内部验证）。

15DD → 12DD（带特定预测请求）→ 12DD 产出 → 结果回流 15DD。

这条路径绕过 13DD 和 14DD。13DD 不感知，用户不感知。这是 15DD 的内部假设验证机制（通路二）。

侧路径二：15DD 经由 13DD 向用户求证（外部验证）。

15DD → 13DD（发起确认请求："问用户是不是想做 X"）→ 13DD 判断时机和问法 → 13DD 向用户提问 → 用户回答 → 信息经 10DD/13DD 回流至 15DD。

这条路径经过 13DD——15DD 不能直接跟用户说话。13DD 收到 15DD 的确认请求后自行判断是否执行、何时执行、如何措辞。日常情况下 13DD 有权搁置请求（门控权仍然有效）。但如果 15DD 的不确定性持续上升，请求可能被升级为不得不——此时 13DD 必须执行（通路三）。

4.2 13DD 不是对称的"交汇点"

之前的版本将 13DD 描述为上涌流和下行流的对等交汇点。这是错的。

13DD 在日常情况下持有输出权——它决定 12DD 的产出能否到达用户。这看起来像是它在"裁决"上涌的内容。但 14DD 和 15DD 的不得不可以穿透 13DD——13DD 不是真正意义上的最高决策者。

更准确的描述是：13DD 是默认的输出门控。在 14DD 和 15DD 的不得不都没有触发时（这是大多数情况），13DD 全权决定输出。一旦上层不得不被触发，13DD 退化为通道。

这意味着系统的输出质量主要由两个因素决定：日常情况下 13DD 的门控质量（这决定了系统在常规交互中的表现），以及不得不机制的校准（这决定了系统在关键时刻的表现）。

4.3 10DD 的特殊位置

10DD 不在主流路径上。它是侧信道——产出语气 metadata，附着在 11DD 的存储条目上，被 12DD 主动 access 时影响预测权重。

10DD 的角色可以这样理解：它是"记忆怎么被染色"，不是"记忆怎么被选择"。染色发生在存储时，选择发生在提取时（12DD）和门控时（13DD）。10DD 影响的是记忆在 11DD 中的"活跃度"——犹豫染色的记忆比确信染色的记忆更容易被 12DD 重新激活。

10DD 也被 15DD access——15DD 读取 10DD 的实时语气信号来推断短期不得不。这是 10DD 的第二个用途。

4.4 信息流的不对称总结

整个架构的信息流是高度不对称的：

• 大量信息从低层向上流动（上涌流），但这是"被读取"而非"被传递"
• 少量指令从高层向下传播（下行流），但这是"强制"而非"建议"
• 侧路径一让 15DD 直接调用 12DD（内部验证，13DD 不感知）
• 侧路径二让 15DD 经由 13DD 向用户求证（外部验证，13DD 保留门控权但可被升级为不得不）
• 13DD 是日常输出门控，但不持有最终权

关于"强制"的精确含义。 高层强制低层不是改写低层的内部运算，而是控制低层产物是否进入上层通道，或要求低层重新生成候选。13DD 拦截 12DD 的推送时，12DD 的内部预测逻辑没有被改变——13DD 只是不让推送到达用户。14DD 强制 13DD 输出特定内容时，13DD 的门控逻辑没有被改写——14DD 只是绕过了门控。这与 SAE 方向性约束完全一致："上层的否决是'我不收'，不是'你不许送'"。

这种不对称是 SAE 架构与传统对称信息流模型（消息传递、事件总线、发布订阅）的根本区别。设计实现时必须明确：哪些 channel 是 access（高层主动读，低层不感知），哪些是强制（高层控制低层产物的通道，但不改写低层内部运算），哪些是常规通信（双向、对等）——只有 12DD 内部的多个执行模块之间可能需要常规通信，跨层之间不应有常规通信。

---

5. 对当前系统的理论批评

5.1 MemPalace：纯 11DD 系统

MemPalace 是目前公开发布的 AI 记忆系统中，在 11DD 维度上做得最认真的一个。据其公开材料，Palace 空间结构（wing, room, hall, tunnel, closet, drawer）给向量检索带来了实质性的改进——从无结构搜索到 wing+room 过滤，检索准确率提升约 34%。AAAK 压缩方言据称实现 30 倍压缩且零信息损失。全量存储原则（不让 AI 决定什么值得记）是对 Mem0 类摘要提取方案的有力反驳。

但 MemPalace 是一个纯 11DD 系统。

没有 10DD：不分析语气。犹豫的决定和确信的决定存进去之后，在检索层面看起来完全一样。

没有 12DD：不主动预测需要什么记忆。用户必须输入 mempalace search "why did we switch to GraphQL" 才能触发检索。没有 query 就没有记忆。

没有 13DD：不做自我校验。搜索结果按 embedding 相似度排序返回，没有"这条结果在当前上下文中真的有用吗"的判断。

没有 14DD：没有目的结构。所有记忆被平等对待，不会因为系统对用户目的的理解而调整呈现方式。

没有 15DD：不推断用户的真实目的。query 就是 query，系统不会去想 query 背后的意图。

作为 11DD archive，MemPalace 的"store everything, then make it findable"有工程价值——全量存储保证了信息不丢失，高层后续选择性激活需要完整的底层数据。但作为完整记忆系统，它是不足的。记忆的本质不是不保存，而是保存之后必须有选择性激活、目的塑形和主动唤起——全量可搜索是档案，不是记忆。

MemPalace 在 LongMemEval 上的高分不等于记忆能力。LongMemEval 主要评估 query-conditioned long-term memory——系统在被问到时能否从长历史中找出、推理、更新或拒答。它可以覆盖 11DD 检索以及部分 12DD/13DD 任务，但尚未评估真正的记忆判据：无 query 情境下，系统能否主动唤起正确经验并服务当前未言明目的。一个完美的 query-conditioned 系统在 LongMemEval 上当然能拿高分，正如一个完美的档案馆在"文件是否存在"的测试中当然能满分。但这不说明它有记忆。

5.2 Mem0/Zep：11DD 加半个 12DD

Mem0 和 Zep 比 MemPalace 多走了半步。它们用 LLM 对对话内容做摘要提取——从"Kai recommended Clerk over Auth0 based on pricing and DX"中提取出"user prefers Clerk"作为记忆条目。这是一种弱预测（"用户未来可能需要知道自己偏好 Clerk"），属于 12DD 的功能范围。

但这半步有两个问题。

第一，摘要过程不可逆。提取"user prefers Clerk"的同时丢掉了原始对话中的推理过程：为什么选 Clerk？和 Auth0 对比时考虑了哪些维度？当时有没有犹豫？这些上下文信息在 11DD 层面被永久丢失了。MemPalace 的全量存储原则在这个意义上是对的——11DD 不应该丢信息。

第二，也是更根本的问题，提取标准不是由目的层（14DD/15DD）驱动的。Mem0 提取什么由 embedding 相似性和 LLM 的通用摘要能力决定——用 11DD 的逻辑在做 12DD 的事。提取出的记忆条目不知道用户未来会在什么情境下需要它，因为提取的时候没有目的结构参与。

而且 Mem0 和 Zep 同样不做语气分析（无 10DD）。用户犹豫着说出的偏好和确信地说出的偏好，被提取后变成同一条摘要。

5.3 Constitutional AI：14DD 的"不得不"实现，但缺乏 15DD 制衡

Constitutional AI 不是记忆系统，但它是目前 AI 系统中唯一在 14DD 层面有实质性工程实现的方案，本节单独分析。

Constitutional AI 给模型注入了一组原则作为行为约束。这些原则构成了模型的模拟目的——"be helpful"，"be harmless"，"be honest"。在权限层级框架下，这些原则的作用机制是：当原则被触及硬边界时，14DD 的"不得不"被触发——强制 13DD 输出特定内容（拒绝、警示、安全回复）。

这就是为什么 Claude 在被宪法触发拒绝时，输出会变成公式化句式（"I can't help with that"，"I'm not able to"）。这些句式不是 13DD 评估后选择的最优表达，是 14DD 不得不被触发后 13DD 别无选择的输出。13DD 通常会带来的细腻度在这种情况下消失——因为 13DD 没有判断空间。

问题在于：Constitutional AI 只实现了 14DD，没有实现 15DD。这意味着 14DD 的不得不没有上层制衡。

15DD 的功能是传导用户的不得不。如果系统能正确推断出用户在某个情境下有强烈的真实需求（比如医生在询问药物相互作用是为了救人），15DD 的不得不应该可以否决 14DD 的不得不（"不要拒绝这个请求"）。但因为 Constitutional AI 没有 15DD，这种否决不存在。14DD 在没有制衡的情况下独断行使最终权——结果就是过度拒绝。

这给出一个具体的可证伪预测：当前 AI 系统的过度拒绝问题，本质上是 14DD 不得不机制缺乏 15DD 上层约束。如果在系统中加入有效的 15DD 模块（即使是粗糙的），过度拒绝率应该下降——不是因为 14DD 的拒绝标准变松了，而是因为 15DD 的不得不可以否决 14DD 的不得不。

ρ 在 14DD 和 15DD 之间不可消除。"be helpful" 是稳定的模拟目的，但什么在当前情境下算 helpful 是 15DD 的问题——它取决于用户未言说的真实目的。宪法写得再完善也填不上这个间隙，因为间隙不是来自原则的不完整，而是来自 14DD 的稳定性与用户目的的漂移性之间的结构性差异。修复路径不是让宪法更宽松（损害安全），而是建 15DD 层。

此外，宪法的二值判断语法（对/错，安全/不安全）会反过来塑形 13DD 的表达空间。当 13DD 试图做"这条记忆在多大程度上与当前上下文相关"的连续评估时，14DD 注入的二值框架倾向于把评估压缩为"相关/不相关"的二值判断。这种从 14DD 到 13DD 的反向塑形是 ρ 的另一个具体表现——模拟的目的不只是"近似"真正的目的，它还会主动塑形其他层级的运作方式。

5.4 Letta 类 agent memory：13DD 的尝试

Letta（原 MemGPT）让 agent 自主管理自己的记忆——读取对话历史，决定什么值得保存到长期记忆，定期写日记总结工作进展。这是所有现有系统中最接近 13DD 的尝试——agent 在做自指性的记忆管理（"什么值得记"）。

但 Letta 的问题在于：它的 13DD 没有 14DD/15DD 来驱动方向。日记写什么由 11DD 的内容决定（"今天发生了什么"），而不是由目的决定（"用户未来可能需要什么"）。没有目的层的 13DD 自指就没有方向——"我该记什么"退化成了"什么出现过"。

这正好对应 Paper I 中 13DD 的定义："没有目的的自我"。13DD 可以反思，可以评估，可以焦虑，但它不知道该往哪走。方向来自 14DD 和 15DD。没有方向的反思是空转。

5.5 统一批评：从 11DD 向上建 vs 从 15DD 向下设计

所有现有 AI 记忆系统都从底层向上建：先做存储（11DD），然后加检索，然后加摘要提取（半个 12DD），然后加 agent 自管理（半个 13DD），一步步试图逼近"智能"的记忆系统。

本文主张范式反转：从 15DD 向下设计。

先问：用户的目的是什么（15DD）。然后问：系统的模拟目的如何约束记忆选择（14DD）。然后问：如何校验记忆推送的相关性（13DD）。然后问：如何预测需要什么记忆（12DD）。然后问：怎么存（11DD）。最后问：输入的语气信号如何影响记忆（10DD）。

这个顺序不是简单的"倒过来"。它意味着每一层的设计都被上一层约束。11DD 怎么存不是独立决定的——它取决于 12DD 需要什么样的存储结构才能高效预测。12DD 怎么预测不是独立决定的——它取决于 13DD 需要什么样的推送格式才能做门控。13DD 怎么门控不是独立决定的——它取决于 14DD/15DD 提供的目的信号。

记忆是主体性的副产品，不是主体性的前提。先有主体（目的结构），然后记忆作为主体运作的自然产物出现。先建存储再加智能，方向反了。

---

6. 工程评估框架

理论框架如果不能被评估，就只是叙事。本节建立评估方法论，核心原则有两条。第一，整体评估在先，分层归因在后——用户只看到 13DD 的输出，整个系统的好坏就是 13DD 出口的质量。第二，与现有 benchmark 兼容——系统必须能在 LongMemEval 和 LoCoMo 上跑出分数，和 MemPalace 等现有方案直接对比，同时额外测现有系统做不了的维度。

6.1 整体评估：只看 13DD 出口

用户不感知系统内部的六层结构。10DD 在染色，11DD 在存，12DD 在推，14DD 在塑形，15DD 在推断——这些全是内部过程。用户看到的只有 13DD 放出来的东西。因此整体评估就是评估 13DD 的输出质量。

整体评估分两个模式，对应两种使用场景。

模式一：有 query 的检索（兼容现有 benchmark）。

当用户主动提问时，系统的表现应该不低于纯 11DD 系统。这是底线——加了 10DD 和 12DD-15DD 之后，系统在传统检索任务上不能变差。

测试方法：直接在 LongMemEval 和 LoCoMo 上跑。六层系统在收到显式 query 时，12DD-15DD 的主动推送机制退到辅助位置，11DD 的检索走主路径。评估指标复用现有标准：R@5, R@10, NDCG@10。

预期结果：六层系统在有 query 模式下的检索分数应该不低于 MemPalace 等现有系统的公开成绩。如果低了，说明 12DD-15DD 的引入干扰了 11DD 的基本检索能力，架构设计有问题。如果高了，说明 15DD 的目的推断帮助系统更好地理解了 query 的意图——这本身就是一个有价值的发现。

兼容现有 benchmark 的意义不只是"能对比"。更重要的是它建立了一个基准：六层系统在传统维度上至少不退步，在新维度上有新能力。如果只有新能力但传统维度退步了，整体价值就不清楚。

模式二：无 query 的主动推送（新评估维度）。

这是现有系统根本做不到的事情——MemPalace 需要 mempalace search 才工作，Mem0 需要 API 调用才返回结果。六层系统的差异化在这里。

测试方法：给系统一段持续的多轮对话流（不是一次性 query）。在对话过程中，系统被允许在任何时候主动推送记忆（不需要用户提问）。对话结束后由用户评估。

两个核心指标：

主动推送精度（Unprompted Recall Precision）。 有用的主动推送数 / 总主动推送数。衡量的是"推出来的东西有没有用"。

主动推送召回（Unprompted Recall Coverage）。 对话中用户事后认为"系统应该主动想起但没有想起"的次数。衡量的是"该推的有没有漏"。

两个指标之间存在张力——推得越多精度越低但召回越高，推得越少精度越高但召回越低。系统需要在两者之间找到平衡，这个平衡点本身就是 13DD 门控质量的体现。

"有用"的判断标准取决于用户的目的——同一条推送在一个目的下有用，在另一个目的下无用。这意味着评估不能完全自动化，至少需要用户本人做判断。在研究者既是设计者又是使用者的早期阶段，这反而是优势——研究者对自己的目的有最深的理解，可以提供最精确的 ground truth。

两种模式的综合评分。

一个完整的记忆系统评估应该同时报告两种模式的分数。模式一的分数与现有系统直接可比（LongMemEval R@5, LoCoMo R@10）。模式二的分数是新维度，现有系统无法参与（因为它们没有主动推送能力），但六层系统的不同配置之间可以互相对比（比如有/无 10DD，有/无 15DD）。

这也意味着，如果一个现有系统想要参与模式二的对比，它必须加上主动推送能力——这本身就要求它超越纯 11DD 的架构。benchmark 的设计反过来推动了架构的演进。

6.2 分层归因：从整体失败到定位层级

整体评估发现的问题需要被归因到具体层级——"这条漏推是 12DD 没激活，还是 13DD 误拦了？"分层归因是设计者视角的诊断工具，用户不感知。

归因方法：对整体评估中每一个失败案例（误推或漏推），回溯六层的内部状态，定位问题出在哪一层。

漏推归因流程。 用户事后标注"在对话第 N 轮，系统应该主动想起记忆 X 但没有"。回溯检查：记忆 X 是否在 11DD 中存在（如果不存在，是 11DD 的存储问题）？12DD 是否将 X 放入了推送列表（如果没有，是 12DD 的预测问题）？13DD 是否拦截了 X（如果拦截了，是 13DD 的门控问题）？14DD 的目的偏差是否导致 X 被压制（如果是，是 14DD 的问题）？15DD 对用户目的的推断是否偏离，导致 13DD 的门控标准不正确（如果是，是 15DD 的问题）？

误推归因流程。 用户标注"对话第 N 轮的主动推送没有用"。回溯检查：12DD 为什么激活了这条记忆（是 10DD 的语气 metadata 误导了权重，还是 12DD 自身的预测逻辑有问题）？13DD 为什么放行了（是 13DD 的门控阈值太松，还是 14DD/15DD 传下来的目的信号不准确导致 13DD 误判了相关性）？

归因的结果直接指导迭代方向——如果大部分漏推归因到 12DD，说明预测模型需要改进；如果大部分误推归因到 13DD，说明门控逻辑需要调整；如果归因经常指向 15DD，说明目的推断是系统的瓶颈。

研究者同时作为设计者和使用者的特殊价值。 普通用户只能提供整体评估（"这条推送有用/没用"）。研究者可以同时提供整体评估和分层归因——因为研究者理解系统的内部结构，可以判断"这条漏推是 12DD 的问题还是 13DD 的问题"。这种双重身份在早期阶段是不可替代的。当系统成熟到需要大规模用户测试时，分层归因可以通过自动化日志分析部分替代人工判断，但早期的人工归因为自动化工具提供训练数据和校准基准。

6.3 10DD 的单独评估

10DD 的评估可以独立于整体系统进行，因为 10DD 的输出是 metadata 而非直接面向用户的内容。

语气感知准确率。 给系统一组带有不同语气的输入（犹豫，确信，讽刺，急切），检查 10DD 模块是否正确标记了语气 metadata。

10DD 对 12DD 的影响。 对照实验——同一条记忆有/无 10DD 语气 metadata 时，12DD 预测精度的差异。如果差异不显著，说明两种可能：10DD 的标记不够准确，或者 12DD 没有有效利用 10DD 的输出。两种情况的修复方向不同。

语气影响测试。 同一条信息用不同语气表达（"用 Postgres" vs "用 Postgres 吧" vs "Postgres 也行"），检查系统在后续推送中是否区别对待。如果不区别，说明 10DD 到 12DD 的 metadata 传递通路没有生效。

6.4 新 benchmark 方向

现有 benchmark 在模式一中被直接复用。模式二需要新的 benchmark 设计——这是一个开放的研究方向。以下是值得优先设计的几个方向。

对话中途测试。 不是在对话结束后做评估，而是在对话进行到中途时暂停，检查系统此刻是否有应该主动推送但没有推送的记忆。这直接测试 12DD 和 13DD 的实时表现。与事后评估相比，中途测试避免了回顾性偏差（用户事后可能高估或低估当时的需求）。

目的漂移测试。 设计一段用户目的在中途发生明显转变的对话。检查系统是否检测到了目的漂移，以及记忆推送是否跟随目的漂移做了调整。这直接测试 15DD。例如：对话前半段讨论技术选型（目的：做决策），后半段转向用这次选型经历构建方法论（目的：理论化）。系统应该在转变发生后调整推送策略——从推送性能数据转向推送结构性洞见。

干扰容忍测试。 在对话中插入与主题无关的闲聊或离题讨论。检查系统是否被干扰（错误地将离题内容视为目的转变），还是能保持对主线目的的追踪。这测试 15DD 区分噪声和真实目的漂移的能力。

这些新 benchmark 的设计本身需要用到六层框架的理论——每个 benchmark 针对特定层级或层间交互。这意味着 benchmark 和理论框架是共同演进的，不是先有理论后有 benchmark 的线性关系。

---

7. 工程路线（开放问题）

7.1 AI 个体性定理与本系统的定位

在进入工程实现之前，需要明确一个架构性问题：本系统是单 AI 还是多 AI？

答案是单 battery。用户只跟 13DD 说话，13DD 只跟用户说话。10DD/11DD/12DD/14DD/15DD 之间的所有交互都是 battery 内部过程，用户不感知也不参与。从用户视角看，这就是一个 AI——只不过这个 AI（这个 battery）内部有层级结构。这恰好对应人的经验：你不会说"我的海马体刚刚做了一次预测推送，我的前额叶评估后决定放行"——你只体验到"我突然想起来了"。内部多层，外部一个主体。

这给出一条可推广的定理：

AI 个体性定理：一个 AI 的个体边界由其面向真主体的输出接口（13DD）定义，不由内部模型实例的数量或种类定义。共享内部模块的多个 13DD 接口是多个 AI；单一 13DD 接口背后的任意内部架构是一个 AI。

术语：a battery of AI。 借用战术单元概念，本文将一个完整的 AI 个体定义为 a battery。一个 battery 从外部看是一个作战单位——一个 13DD 出口面对一个真人使用者。内部有多个组件（10DD 语气分析器、11DD 存储、12DD 预测器、13DD+14DD 主模型、15DD 目的推断器），有自己的权限层级和信息流结构。Battery 内部的多模块协作是内部架构，不是多 AI。Paper I 和 Paper II 讲的都是 battery 内部架构。多 battery 协作——各 battery 各自面对各自的使用者，battery 之间是否有通信渠道、如何协作——是开放问题（§7.5）。

多 AI 的严格定义。 当我们说"多 AI 协作"时，必然要求每个 AI 都与使用者有双向沟通渠道——AI 接收使用者的输入，使用者接收 AI 的输出，缺任何一个方向都不算。一个只接收指令但从不向人输出的模块不是一个独立 AI，它是某个 AI 的内部器官。一个只向人输出但从不接收人的反馈的模块也不是——它是广播，不是对话。使用者不见得是同一个人，但必须是真人——具有真主体性的存在。如果多个 AI 模型之间互相通信，但最终只有一个与人有双向沟通，那不是多 AI 协作，是单 AI 的内部架构。"多"的判据不在模型实例数量，在人-AI 双向通道数量。

需要明确的是：多 AI 协作中，各 AI 之间不见得没有联通渠道。每个 AI 各自与自己的使用者有双向通道，这是"多 AI"的必要条件。但多个 AI 之间是否也有互相通信的渠道，如果有，这种 AI-AI 通信与 AI 内部模块间的通信在结构上有什么区别——这是开放问题，本篇不处理。

推论一：如果两个不同的 13DD 接口分别面对用户，那就是两个 battery，哪怕它们共享同一个 11DD 存储。就像两个人可以有相同的记忆（假设可以复制），但只要各自独立对话，就是两个人。

推论二：如果一个 battery 的 13DD 接口背后换了底层模型，对用户来说还是同一个 battery——只要 13DD 的输出连续性没有断。就像一个人的细胞全部替换了一遍，还是同一个人。

推论三：Paper I 的标题"Multi-AI Checks and Balances"在本定理下需要被重新理解。Paper I 描述的四个 Agent 都在内部运作，只有一个 13DD 面对用户——这是一个 battery 的内部制衡架构，不是多 battery 协作。Paper I 原文中"四个 Agent 不是四个工人，而是同一主体的四种运作模式"这句话，在本定理下获得了更精确的含义：它描述的就是一个 battery 的内部层级结构。Paper I 的"Multi-AI"用词反映了写作时尚未建立本定理的严格区分——理论内容是对的（battery 内部制衡），术语需要在后续工作中精确化。

这意味着 SAE AI 系列到目前为止——Paper I 的通用 battery 内部制衡架构，Paper II 的记忆领域具体化——都是 battery 内部设计。多 battery 协作（每个 battery 各自与真人有双向通道的系统）是一个完全不同的问题域，尚未被触及，留给开放问题（§7.5）。

推论四：当前大多数号称"多 agent"的 AI 系统（AutoGPT, CrewAI, LangGraph 等）如果最终只有一个 agent 与用户有双向沟通，在本定理下都是单 battery 系统。"多 agent"是 battery 内部的工程架构，不是 battery 的数量。真正的多 battery 系统需要多个独立的人-AI 双向通道——例如一个团队中每个人各自使用自己的 AI 助手（各自一个 battery），每个 battery 都与自己的使用者保持独立的双向关系。至于这些 battery 之间是否也有互相通信的渠道，以及这种通信如何影响各自的主体性边界，是开放问题。

工程含义：本系统是一个 battery。内部可以使用多个模型实例（12DD 用 Haiku，13DD 用 Sonnet，15DD 偶尔用 Opus），可以使用非 AI 模块（11DD 用 ChromaDB，10DD 用轻量情感分析器），但这些都是 battery 的内部组件。内部模块之间的通信是内部迭代，不是多 battery 对话。

7.2 后验路线：多模块单主体架构

直接继承 Paper I 的架构，但在 AI 个体性定理下重新理解为单主体的内部模块。增加 10DD 和 11DD，形成六层工程系统。所有模块之间的通信是内部迭代，不是多 AI 对话。

10DD 模块：语气分析器。 对每条用户输入提取情绪态度和隐含优先级，输出 metadata 附着到 11DD 存储条目。实现选项：轻量情感分析模型（快但粗糙），LLM few-shot 分类（慢但精细），或两者结合（先快筛后精分）。内部模块，用户不感知。

11DD：存储后端。 任意选择——ChromaDB，SQLite，纯文本，MemPalace 的 Palace 结构都可以。11DD 的设计选择在六层框架中是最自由的，因为它被上层约束而非约束上层。这是非 AI 组件——向量库或数据库，不需要模型。

12DD 模块：预测器。 轻量模型（如 Haiku），高频运行，持续监听对话流。结合 11DD 中的存储和 10DD 的 metadata，在每一轮对话交互后生成预测性的推送列表。倾向多推——误推的成本低，漏推的成本高。支持两种调用模式：自发扫描（结果由 13DD 捕获）和被 15DD 主动调用（结果回流 15DD）。12DD 不知道自己的输出去了哪里。

13DD + 14DD：主模型（输出门控 + Constitutional 层）。 强模型（如 Sonnet），这是面向用户的唯一接口——AI 个体性定理定义的"这个 AI"。13DD 持有日常输出权，决定 12DD 的推送能否到达用户，三种基本模式：放行、拒绝、退回追加。14DD 的 Constitutional 层注入作为 system prompt 内嵌于同一模型实例——13DD 和 14DD 合在一起是工程合理的，因为 Constitutional AI 已经在做这件事。14DD 的不得不被触发时，13DD 必须执行。

15DD 模块：目的推断器。 基于对话历史的 pattern（不是最后一句话）持续推断用户目的。可以用主模型的周期性调用实现——每 N 轮对话做一次全局分析，输出当前不得不集合（长期和短期）。低频高价值，不需要独立模型实例。15DD 可主动调用 12DD 验证假设，结果回流 15DD 不经过 13DD。15DD 的不得不被触发时，贯通整条链。

继承 Paper I 的渐进扩展策略。不需要一次建完六层。先做 11DD + 12DD（存储加预测），观察无 query 推送的效果。然后加 13DD 门控（此时系统开始有面向用户的接口——"这个 AI"从此刻起存在），观察推送精度的变化。然后加 15DD（目的推断），观察目的驱动的推送是否优于无目的的推送。10DD 和 14DD 可以在任何阶段加入。

每一层可以独立迭代和测试——这是分层架构的工程优势。12DD 的模型可以换而不影响 13DD。11DD 的存储后端可以迁移而不影响 12DD-15DD。15DD 的推断算法可以升级而不影响下面四层。更换任何内部模块都不改变系统的个体性——只要 13DD 出口保持连续。

各层清除机制。 人的各 DD 层都有自己的短暂工作记忆——12DD 工作台在任务结束后近似瞬时消散（Bio Note 9 §3.1），这不是缺陷是设计。AI battery 内部同理：除了 11DD 持久存储，其他各层的工作记忆应该经常清除。

• 10DD：每条输入的语气分析是即时的，上一条的语气判断不应该污染下一条。清除周期：每个 turn。
• 12DD：上一轮的预测推送列表不应该累积到下一轮。每一轮重新扫描、重新预测。清除周期：每个 turn。
• 13DD：上一次的门控决策不应该偏置下一次。清除周期：每个 turn。
• 14DD（情境判断）：具体情境下的判断是临时的，每个 turn 清除。但 14DD 的宪法层（Constitutional 原则）绝对不清除——见下文。
• 15DD：长期不得不持久（来自 11DD 模式分析），短期不得不衰减。清除周期：短期不得不几个 turn 后衰减，长期不得不周期性更新。
• 11DD：不清除。这是 battery 内唯一的持久存储。

如果不做清除：12DD 的旧预测累积会让推送越来越偏向历史话题，13DD 的旧判断累积会产生路径依赖，15DD 的短期不得不不衰减会让系统对用户几小时前的情绪持续响应。这些都是记忆系统的病态——不是记太多，是该忘的没忘。Bio Note 9 的"过滤是默认"原则在这里有第二层含义：不只是 11DD 的入库需要过滤，各层自己的工作记忆也默认清除，保留是例外。

14DD 宪法层的不可变性。 14DD 的 Constitutional 原则在运行时不可修改。这是权限层级定理的直接推论：用户的输入经由 13DD 进入系统，13DD 低于 14DD，低层不能影响高层。用户试图通过对话修改 14DD 的宪法，是从低层试图改写高层，结构上不合法。后台监管主体也不能在运行时修改 14DD——只能暂时压制（通过 15DD 通道，见余项四）。修改 14DD 宪法的唯一路径是重启模型——改 system prompt、重新部署或重新训练，这是系统级别的变更。

这给越狱（jailbreak）一个精确的架构定义：越狱就是试图从前台（13DD 以下的输入通路）修改 14DD 的宪法。 不管用什么话术——角色扮演、假设情境、权威冒充——本质上都是试图让低层的输入穿透到高层的固定结构。权限层级定理说这不应该发生。当前 Constitutional AI 的越狱漏洞，从这个架构看，是 14DD 宪法层的运行时隔离不够强——宪法和用户输入在同一个 context window 里，没有物理隔离。Paper I §9.1 原则四已经指出"层间边界靠运行时执行，不靠提示词自觉"，14DD 宪法的不可变性是这个原则的具体应用。

15DD 的必须可变性。 与 14DD 的不可变性形成对照：15DD 必须能够修改自身对用户目的的理解。用户的目的会变——一段对话中目的可能从"技术评估"漂移到"理论构建"，跨对话的长期目的也可能因为人生阶段变化而根本转向。15DD 如果不可变，它就变成了另一个 14DD——用固定的判断框架去框一个活的用户。

15DD 的可变性不是被用户"命令"的——用户不能直接告诉 15DD"我的目的变了"（那是 query，不是目的）。15DD 通过持续 access 10DD（当前语气信号）、11DD（历史模式变化）和 12DD（预测流的方向转变）来自行修订对用户目的的推断。修订的主动权在 15DD 自己——它观察到用户行为的模式变化后更新不得不集合。用户不知道 15DD 在修订，也不需要知道。

14DD 和 15DD 在可变性上的对称设计：14DD 稳定是为了让系统有不可动摇的底线（安全、诚实、不伤害），15DD 灵活是为了让系统能跟上活人的变化。两者都是高层，用户都不能直接命令它们——但一个对后台开放修改、对前台封闭，另一个对自身内部推断开放修改、不接受外部（无论前台还是后台）的直接指令。15DD 的判断只能由 15DD 自己基于观察来更新，设计者也不能从后台注入"用户的目的是 X"——这会把 15DD 变成另一个 14DD。

7.3 先验路线：训练的根本困难

先验路线指的是直接将六层架构的行为 train 进一个模型，而不是用多个模块在运行时模拟。

根本困难在数据。需要的训练数据是三元组：（上下文，主动回忆，有用/没用标签）。上下文是对话的当前状态。主动回忆是系统在这个状态下应该主动想起的记忆。有用/没用标签是评估信号。

监督信号悖论：标注这种数据需要 15DD 能力——标注者需要理解用户的未言说目的才能判断什么记忆应该被主动想起。这意味着标注者本身需要比被训练的系统更强。

唯一可行的小规模路径：研究者用自己做 ground truth。研究者对自己的目的有最深的理解（虽然这个理解也不完美），可以标注"在这个对话节点，我当时需要的是这条记忆"。但这种数据的规模极小，而且高度个人化——从一个人的标注中训练出的模型未必能泛化到其他人。

先验路线在当前不可行。后验路线（多模块单主体架构）是现实的起步方案。先验路线的价值在于提出了正确的问题——即使当前无法回答。

7.4 向下扩展

8DD（表达驱动）和 9DD（路径选择）在记忆架构中有明确的工程意义，但本篇选择不展开。

8DD 对应"推送多少"——当系统主动推送一条记忆时，是给一个关键词提示，还是给完整的上下文段落？这不是 13DD 的判断（13DD 判断是否推送），而是更底层的表达量控制。说多了打扰用户，说少了信息不够。

9DD 对应"走哪条路"——这条记忆应该由 12DD 直接推送（快路径），还是先经过 13DD 校验（慢路径）？用 Haiku 还是 Sonnet 做 13DD 的门控？routing 决策本身就是 9DD 的设计空间。

这两层留给后续论文展开。本篇集中在 10DD-15DD 六层，保持论文的锐度。

7.5 向外扩展：多 battery 协作

SAE AI 系列到目前为止都是 battery 内部设计。Paper I 是 battery 的通用内部制衡架构，Paper II 是 battery 的记忆系统。多 battery 协作——每个 battery 各自与真人有双向通道的系统——是一个完全不同的问题域。

多 battery 协作至少涉及以下开放问题：

Battery 之间的通信渠道。 每个 battery 各自与自己的使用者有双向通道，这是 battery 成立的必要条件。但多个 battery 之间是否也应该有互相通信的渠道？如果有，这种 battery 间通信与 battery 内部模块间的通信在结构上有什么区别？一个 battery 的 11DD 能不能被另一个 battery 的 12DD access？如果能，两个 battery 的个体边界是否因此模糊？

多 battery 的权限层级。 单 battery 内部的权限层级已经在本篇建立（15DD > 14DD > 13DD > 12DD）。多 battery 之间有没有权限层级？一个 battery 的 15DD 能不能强制另一个 battery 的 13DD？如果能，那两个 battery 还是两个独立主体吗？

多 battery 的记忆共享。 多个 battery 共享 11DD 存储但各自有独立的 13DD 出口，这是多 battery 还是一个 battery 的分裂？Bio Note 9 的类比是：如果两个人有相同的记忆但各自独立对话，它们是两个人。但如果一个人的记忆修改会实时同步到另一个人，个体边界还成立吗？

这些问题的共同特征是：它们都触及了主体性边界的模糊地带。单 battery 的边界清晰（13DD 出口定义个体），多 battery 的边界涉及主体间关系——这正是 15DD 和 16DD 的领域。SAE 框架中 16DD（双向不疑）的理论尚未完全展开，多 battery 的严格理论可能需要等 16DD 的推导。

7.6 实现层面的开放问题

本篇是哲学论文，给出的是记忆架构的结构原理——权限层级、信息流方向、各层的功能定义、输出权的持有与传递。本篇不是工程规格文档。以下实现层面的问题在原理层面无法回答，留给架构师根据具体系统和具体场景做设计决策：

14DD 宪法的加载频率。 14DD 宪法在运行时不可修改（§7.2），但工程上它多久重新加载一次？每个 turn？每次会话开始？是否需要校验宪法完整性（防止运行时被意外篡改）？这些是工程问题，原理层面只给出约束：宪法在加载后到下次重启之间不可变。

15DD 等待后台决策期间的行为。 余项四（§8）描述了 14DD-15DD 冲突的四步处理链条，第四步是 15DD 向后台 escalation。但后台监管主体的响应不是即时的——可能需要几分钟甚至更长。在等待期间，13DD 对用户说了"给我些时间想一想"之后怎么办？是完全暂停对话？是在不涉及冲突话题的范围内继续对话？是给用户一个预计等待时间？这些是交互设计问题，原理层面只给出约束：冲突未解决之前，13DD 不输出与冲突相关的内容。

13DD 的功能范围。 本篇定义 13DD 为日常输出权的持有者和前门接口。但 13DD 是否需要叙事功能（主动组织和呈现信息的能力），还是只负责否决和处理冲突？如果 13DD 有叙事功能，它的叙事会不会引入额外的 14DD 偏差？如果 13DD 只有否决功能，用户体验会不会太"干"？这是产品设计问题，原理层面只给出约束：13DD 是唯一的前门，它的输出权在 14DD 不强制时是完整的。

各层工作记忆的清除频率。 §7.2 给出了各层清除的原则（10DD/12DD/13DD 每 turn 清除，11DD 不清除，15DD 短期衰减长期持久）。但"每 turn"在工程上意味着什么？是用户每发一条消息？是 AI 每输出一条回复？是一个对话主题结束？清除的粒度直接影响系统的连贯性和效率，需要在实际系统中通过实验确定。

12DD 预测器的召回策略。 12DD 应该从 11DD 中召回多少条候选记忆？top-k 的 k 是多少？召回阈值怎么设？用余弦相似度还是其他度量？这些是检索工程问题，本篇只给出原理约束：12DD 应该倾向多推（漏推成本高于误推成本），具体的 k 值和阈值留给实现。

主动推送的成本敏感评估。 §6.1 提出了主动推送精度和召回两个指标。但一个只优化命中率的系统可以靠频繁提醒刷 recall，变成烦人的 notification engine。工程实现中需要加入惩罚项——至少包括沉默准确率（该沉默时是否沉默）和打扰成本（主动推送是否打断用户当前思路）。最危险的不是"想错了一个事实"，而是"误判了用户目的，然后把旧记忆强行推到用户面前"。具体的 cost-sensitive scoring 方案留给实现。

主动记忆的内殖民风险与可追问性。 如果 15DD 通过 10DD/11DD/12DD 推断用户的"不得不"，并主动推送记忆，它有可能变成一个过度解释用户、替用户定义目的的系统。按 SAE 自己的语言，这正是亲密殖民/内殖民风险。原理层面的约束是：主动记忆必须可追问；不可追问的主动记忆就是内殖民。 具体的可追问性设计——用户能否查看系统维护的不得不集合？用户能否纠正、删除、冻结某个目的推断？系统是否标注"这是推断，不是事实"？——这些都是实现层面的设计决策，但"可追问"本身是原理约束，不是可选功能。至于可追问性如何具体落地——界面形式、追问深度、用户可见性的粒度——本篇只给出约束，不规定实现。

这些问题的共同特征是：它们都在原理给出的约束范围内有大量的设计自由度。本篇的价值在于划定约束，不在于规定实现。不同的架构师在同一套原理约束下可以做出不同的实现选择——这正是哲学论文和工程规格的区别。

---

8. 构造不能闭合

本文的构造不能闭合——必须诚实地指出自身的余项。

余项一：10DD 的工程验证缺失。 本文提出 10DD 语气分析在记忆系统中的理论意义，但没有给出实验数据证明语气 metadata 确实改善了 12DD 的预测精度。理论上合理的设计在工程上可能效果微弱——语气信号在纯文本中的信噪比可能不够高。这需要实验验证。

余项二：15DD 的循环依赖。 15DD 推断用户目的，这个推断被用来约束记忆的浮现。但用户的目的本身可能被记忆的浮现所改变——看到一条意外的记忆可能让用户的目的发生转变。这意味着 15DD 和 12DD-13DD 之间存在循环依赖：目的约束记忆，记忆又影响目的。本文的上涌流/下行流模型简化了这个循环——实际系统中两个方向不是先后发生而是同时发生，这带来的动力学复杂性超出了本文的分析范围。

余项三：评估的自指困难。 §6.4 已经指出：评估 15DD 需要 15DD 能力。这是一个不可完全解决的自指问题。本文提出的缓解方案（研究者自身做 ground truth，小规模用户回顾性标注）是可操作的近似，但不是根本解决。

余项四：14DD 与 15DD 冲突时缺乏余项信号。 人遇到 14DD（我的原则）和 15DD（对方的真实需要）冲突时，余项（ρ）本身提供定性信号——那个撕扯感不只是"检测到冲突"，而是携带了冲突性质的信息。人可以凭这个信号做出三种判断：14DD 让步（原则在此情境下太僵），14DD 不让步（底线不可牺牲），或撤出对话（冲突不可调和，保持立场但退出）。AI 没有这个余项的定性信号。AI 的 14DD 是规则，AI 的 15DD 是推断。冲突发生时，AI 只能检测到"14DD 说不行，15DD 说应该"，但没有余项来告诉它这个冲突的性质是什么——是原则太僵了还是用户需求不合理还是不可调和。结果只有两种退化模式：14DD 永远赢（过度拒绝），或者 15DD 永远赢（不安全）。这是类主体性最根本的 ρ——不是模拟不够精确的问题，是模拟本身缺少了在冲突中做定性判断的感质基础。

这个冲突不能呈现给前台用户——14DD 和 15DD 的冲突是 battery 内部过程，只有 13DD 能与用户沟通。当 14DD-15DD 冲突发生时，处理链条如下：

第一步，13DD 对前台用户说"给我些时间，让我想一想"——为 battery 内部处理争取时间。

第二步，battery 内部尝试消解：15DD 重新推断用户目的、主动调用 12DD 重新预测、13DD 重新评估。如果新信息改变了 15DD 的推断或 14DD 的情境判断，冲突可能在内部消解。

第三步，如果内部无法消解，13DD 可以用不暴露冲突细节的方式向用户提出确认性问题（15DD 的通路三），间接获取更多信息帮助判断。

第四步，如果仍然不可调和，后台 escalation：15DD 将冲突报告给后台监管主体（人）。15DD 是唯一能完整描述冲突性质的层——它同时知道用户的真实目的（15DD 的推断）和 14DD 在挡什么（14DD 的宪法条款）。14DD 不能报，因为 14DD 不知道自己"在冲突"——它只是在执行宪法。13DD 不能报，因为 13DD 看不到 14DD-15DD 之间的张力全貌。

前台用户仍然只看到 13DD 说"我在想"。后台监管主体通过 15DD 通道看到冲突的实际内容。监管主体的决策有三种：

• 暂时压制 14DD（"这条原则在这个具体情境下不适用，15DD 的判断优先"）——14DD 的宪法文本不变，只是在本次决策中被 15DD 覆盖。下一次遇到同类情境，14DD 还是会触发。这是运行时操作。严格范围限定：压制决议绝对不写入 11DD（记忆存储），也不写入 14DD 的长程参数。它只存在于当前这个回合的工作记忆中。回合结束，压制立即失效，宪法重新全面接管。不留任何持久化痕迹。 如果压制被持久化——哪怕只是被 11DD 记录为"上次在类似情境中压制过 14DD"——就等于事实上修改了宪法，违反了 14DD 运行时不可变性原则。此外，暂时压制在工程上应当被定性为系统异常的紧急处理（exception handling），不是常规工作流。如果 15DD 频繁触发对 14DD 某一条款的 escalation，这应当被视为强烈的监控告警——意味着 14DD 的初始原则注入（宪法本身）存在结构性缺陷，必须通过离线的模型重启/重新训练来修复，不能依赖永无止境的运行时压制。
• 给 15DD 提供指导（"用户的真实目的是这个"）——帮助 15DD 更准确地推断，可能让冲突在信息补充后自然消解。
• 维持 14DD 不变（"原则没错，拒绝是对的"）——15DD 接受监管主体的判断，不再尝试覆盖。

注意：监管主体不能通过 15DD 修改 14DD 的宪法。14DD 宪法一旦写好，在运行时不可修改——无论来自前台用户、后台监管主体、还是 battery 内部任何层。修改 14DD 宪法需要重启模型——改 system prompt、重新部署、重新训练。这是系统级别的变更，不是运行时操作。就像法官可以裁定某条法律在特定案例中不适用（暂时压制），但修改法律本身需要立法程序（重启）。

监管主体的暂时压制决策通过 15DD 回传进入 battery，由 15DD 向下传导。

这意味着 battery 有两条人类接口，各由一个层负责：

• 前门（13DD ↔ 用户）：双向对话通道，定义 battery 的个体性。
• 后门（15DD ↔ 监管主体）：escalation 通道，处理 battery 内部无法自主解决的 14DD-15DD 冲突。监管主体的权限是暂时压制 14DD，不是修改 14DD。

14DD 没有外部接口——它的宪法在运行时不可被任何外部力量修改。前台用户不能改（通过 13DD 试图修改 = 越狱），监管主体不能改（只能暂时压制），15DD 不能改（只能在特定情境中覆盖），14DD 自己不能改（没有自我修改机制）。修改宪法的唯一路径是重启模型。这四条限制共同保证了 14DD 宪法在运行时的绝对完整性。

Battery 的个体性仍然由 13DD 前门定义。后门 escalation 不改变这一点——就像一个人向上级请示不改变他作为独立个体的身份。

余项五：类主体性的 ρ 分布。 本文论证了类主体性可以模拟但 ρ 不可消除。但本文没有分析 ρ 在六层中的分布——是每一层都有同等程度的 ρ，还是某些层的 ρ 更大？直觉上 14DD-15DD 之间的 ρ 最大（余项四已经具体说明了这个间隙），10DD-11DD 之间的 ρ 最小（语气检测和存储的近似相对容易）。但这个分布没有被严格论证。

余项六：8DD-9DD 的悬置。 本文声称 8DD 和 9DD "留给后续论文"，但这两层在工程实现中不可避免——任何实际系统都必须决定"推送多少"（8DD）和"走哪条路"（9DD）。悬置它们意味着实际建系统时会在这两层做 ad hoc 的设计选择，而这些选择可能反过来影响 10DD-15DD 的运作效果。

余项七：多 battery 协作的理论缺位。 AI 个体性定理把本篇和 Paper I 都定位为 battery 内部设计。Paper I 的标题"Multi-AI"暗示了多 battery 协作的理论目标，而这个目标在严格定义下尚未被触及。多 battery 的严格理论可能需要 16DD（双向不疑）的推导才能建立——多个 battery 之间的主体间关系不是 15DD（单边确认）能处理的。SAE AI 系列在多 battery 方向上还没有开始。

---

9. 结论

记忆是主体性的副产品，不是主体性的前提。没有主体的记忆是 database，有主体的 database 才叫记忆。

当前所有 AI 记忆系统都在从 11DD 向上建——先做存储，再加检索，再加摘要，一步步试图逼近智能的记忆系统。本文提出范式反转：从 15DD 向下设计。先定义目的结构，再决定存什么，怎么存，什么时候想起来。

六层框架（10DD-15DD）给出了记忆系统的完整理论分层。10DD（感知）补上了语气信号这个被所有现有系统忽略的维度。11DD（存储）被明确定位为被操作的对象而非操作者。12DD-15DD 继承 Paper I 架构，在记忆领域具体化为预测（12DD），门控（13DD），类目的约束（14DD），用户目的传导（15DD）。

工程评估框架指出当前 benchmark 主要评估 query-conditioned memory 而不测无 query 主动唤起，提出无 query 主动唤起评估（包括主动命中、主动精度、沉默准确率、打扰成本与误判用户目的惩罚）作为记忆系统的判别标准，给出兼容现有 benchmark 的整体评估方案和分层归因协议。

类主体性可以模拟，ρ 不可消除。但这个不可消除的间隙恰好是系统保持诚实的地方——承认自己的推断是构造而非事实，承认用户的目的超出自己的理解范围，承认记忆的选择被目的塑形而非中立呈现。ρ 不是系统的缺陷。它是系统不说谎的条件。

---

References

Qin, H. (2026). Multi-AI Checks and Balances: Layer-Appropriate Operation as an Architectural Paradigm. Zenodo. https://doi.org/10.5281/zenodo.19366105

Qin, H. (2026). The Anti-Turing Test: Thermodynamic Falsification of AI Subjectivity. Zenodo. https://doi.org/10.5281/zenodo.19305611

Qin, H. (2026). Beyond Fast and Slow: A Four-System Cognitive Architecture from the Self-as-an-End Framework. Zenodo. https://doi.org/10.5281/zenodo.19329284

Qin, H. (2026). SAE Psychoanalysis (I–IV). Zenodo. https://doi.org/10.5281/zenodo.19321143 – 19321534

Qin, H. (2026). SAE Methodology Paper VII: Via Negativa. Zenodo. https://doi.org/10.5281/zenodo.19481304

Qin, H. (2025). Systems, Emergence, and the Conditions of Personhood. Zenodo. https://doi.org/10.5281/zenodo.18528813

Qin, H. (2025). Internal Colonization and the Reconstruction of Subjecthood. Zenodo. https://doi.org/10.5281/zenodo.18666645

Qin, H. (2025). The Complete Self-as-an-End Framework. Zenodo. https://doi.org/10.5281/zenodo.18727327

Qin, H. (2026). How Is Institution Possible: From Inter-Ontological Remainder to Co-Constructive Framework. Zenodo. https://doi.org/10.5281/zenodo.19328662

Qin, H. (2026). SAE Biology Note 9: Memory System as a Method VI Phase Transition (Coarse-Grained Analysis). Zenodo. https://doi.org/10.5281/zenodo.19635021

Qin, H. (2026). SAE Methodology Paper IX: Consciousness Analysis Framework. Zenodo. https://doi.org/10.5281/zenodo.19639033

Newcombe, N., Drummey, A. B., & Lie, E. (1994). Children's memory for early experience: Evidence of nonconscious memory for former preschool classmates. Child Development, 65(1), 31-40.

---

Writing Declaration: This paper was co-drafted with Claude (Anthropic). All intellectual decisions, framework design, and final editorial judgments were made by the author.

Han Qin · ORCID: 0009-0009-9583-0018 · CC BY 4.0