Chapter 3 — The Cognitive-Science and Computer-Science Tradition
3.0 The same answer, reached from the other side
Chapter 2 traced a philosophical argument: thinkers who built nothing were pushed, objection by objection, toward identity as persistent, inspectable structure. The objection that this is only philosophy agreeing with philosophy is answered here. The same structures ENGRAM is built from — provenance, retraction with cascade, graded confidence, consolidation, designed forgetting, identity as accumulated record — are each precedented, from their own evidence and toward their own goals, in two technical literatures that mostly do not cite each other and neither of which cites us: the cognitive science of human memory, and the computer science of belief maintenance and agent memory. We do not call this independent discovery. The fields share a common root in human cognition — cognitive science studies it, logic and belief revision formalize it, agent-memory work borrows its analogies — and ENGRAM itself was designed by introspecting one human's reasoning. What is independent is narrower: ENGRAM's design was arrived at without its designers having studied these results.
This convergence is the chapter's load-bearing point. ENGRAM has no uniquely novel component; every mechanism is precedented somewhere. Read as a novelty claim, that is a weakness; read as evidence, it is the opposite. When fields reasoning from different evidence toward different ends arrive at the same structure, and a memory built by introspection arrives there too, the structure is likely to track something real about how a knowing mind reasons and revises. The agreement is not a coincidence. Human cognition is the most successful reasoning we know of; ENGRAM's substrate, a language-trained model, inherits that cognition through human language; and its designer introspected the same cognition to build it. Convergence with the fields that study human reasoning is therefore to be expected: empirical science describes the structure from the outside, ENGRAM instantiates it from the inside. Whether the structure is specific to human minds or universal to any reasoning system we leave open; ENGRAM on a non-human substrate is suggestive, but it was designed from human introspection, and we claim no proof of substrate independence. ENGRAM's distinctiveness is accordingly narrower: the integration (no prior system unifies all of these mechanisms in one substrate) and the purpose (an epistemic identity that fails loudly), which follows from the motivating premise of §1.1 — an LLM's hallucination is structural, so reliability must come from external calibration rather than elimination.
The chapter has two strands. Each section names a tradition, states the ENGRAM mechanism it precedes, and points to the Chapter 4 section that builds that mechanism.
Strand I — Cognitive science of memory
3.1 Complementary Learning Systems → continuity across compaction
Complementary Learning Systems theory (Kumaran et al., 2016; McClelland et al., 1995) holds that durable memory needs two coupled subsystems: a fast-learning, instance-specific store (hippocampal) and a slow-integrating, structure-extracting store (neocortical), with consolidation moving information between them. ENGRAM realizes this division symbolically: fast, instance-specific observations against slow, structure-extracting derivations and cornerstones, with its consolidation routines as the transfer process: a nap at the end of each burst (one stretch of work between two context compactions), which writes what the burst produced into nodes before the context is compacted; and a once-daily sleep in two phases, a review that files what the day left unfiled, then a dream phase in which scanner sub-agents (dream fairies) read the graph for contradictions, stale nodes and lesson candidates, and a consolidating sub-agent (the dream master) applies the results and advances the forgetting curve.
→ Ch4 §4.8 (continuity across compaction): CLS is the cognitive-science precedent for the two-routine consolidation that carries operational continuity across compaction.
3.2 Tulving → the two registers of retrieval
Tulving distinguished autonoetic remembering (awareness of an event "as a veridical part of his own past existence") from noetic knowing (awareness of objects and events "in the absence of these objects and events") (Tulving, 1985, p. 3). Tulving and Thomson's encoding specificity principle adds the cue side: "specific encoding operations performed on what is perceived determine what is stored, and what is stored determines what retrieval cues are effective in providing access to what is stored" (Tulving & Thomson, 1973, p. 369). ENGRAM's retrieval has exactly two registers, by design: a lossy auto-surface (noetic, hint-level: "this sounds familiar") and deliberate recall (a query whose hits the agent can open in full: "I need to know exactly"). The lossy auto-surface is what prompts the agent to invoke deliberate recall. Reading a node is noetic, not autonoetic: ENGRAM provides remembered knowledge without phenomenal re-living, the boundary Ch2 §2.3 also marks.
→ Ch4 §4.6 (auto-surface and deliberate recall): Tulving's two registers are the precedent for the auto-surface, and encoding specificity (Tulving & Thomson, 1973, p. 369) for its cue-dependence; the auto-surface binds Condition 3.
3.3 Rational analysis of memory → designed forgetting
Anderson and Schooler (1991) showed that human forgetting is not a defect but an adaptation: what stays available tracks the need probability — "the probability that we will need a particular memory trace now" — estimated from environmental frequency, recency, and spacing. Ebbinghaus (1885) measured the original forgetting curve (Ebbinghaus, 1885/1913, sec. 29, p. 76). Wixted and Ebbesen found a simple power function described forgetting best in their experiments, and that a reanalysis of Ebbinghaus's savings data declines as a power function too (Wixted & Ebbesen, 1991); Wixted's later review finds the exponential describes forgetting poorly (Wixted, 2004, p. 243). Whether individual memories decay exponentially, with the power law an artifact of averaging, is contested (Anderson et al., 1999, p. 1136). ENGRAM's per-node exponential is a design choice for stationary recency ordering, not a claim about human forgetting.
ENGRAM forgets by design: recall importance decays, maintaining a bounded queryable set and a faded set (retained, not destroyed), with recall importance governing what stays queryable (Ch4 §4.5) and a usage-learned utility score shaping how results are ranked (Appendix C). This is the need-probability principle instantiated: what the agent is most likely to need stays queryable; the rest fades.
→ Ch4 §4.5 (forgetting): the rational-analysis tradition is the cognitive-science anchor for ENGRAM's queryable/faded sets; the exponential kernel is ENGRAM's own design choice.
Strand II — Computer science and formal epistemics
3.4 Truth Maintenance Systems → provenance + retraction cascade
Truth Maintenance Systems attach justifications to beliefs. In Doyle's
justification-based TMS, each belief is in or out according to its justifications, and
when a justification no longer holds, the status of the beliefs that depend on it is
recomputed along the dependency links (Doyle, 1979/1987); contradictions are removed by
dependency-directed backtracking, which adds justifications (de Kleer, 1986, p. 129). de Kleer's assumption-based TMS instead labels each datum
with the assumption sets under which it holds, and so avoids most backtracking and all
retraction (de Kleer, 1986, pp. 127, 129). Doyle's mechanism is ENGRAM's derives_from
provenance and retraction cascade, decades earlier and in a different field. Of the
2023–2026 LLM-agent papers we examined, only Wray, Kirk and Laird cite Doyle, and only
to describe Soar's truth maintenance (Wray et al., 2025, sec. 2); none of the memory systems we
surveyed (Appendix B) automatically propagates retraction along dependencies (Kumiho's
impact analysis is advisory). ENGRAM implements TMS-style dependency propagation in typed LLM-agent
memory, and names the lineage it implements.
→ Ch4 §4.2 + §4.4 (provenance primitives; retraction cascade → Condition 3).
3.5 Subjective Logic → confidence propagation, with a divergence
Jøsang's (2013) Subjective Logic is an algebra of graded opinions. An opinion about a claim
is a tuple (belief, disbelief, uncertainty, base rate) with b+d+u=1: how far the
evidence supports the claim, how far it counts against it, and how much room is left.
Operators combine opinions from several sources. ENGRAM's confidence belongs to this
family, and its propagation is not novel. ENGRAM's single number is closely
analogous to SL's belief mass b with the disbelief mass set to zero: it says how far
the evidence supports the claim as stated, and 1−b is the room left.
The divergence: ENGRAM has no disbelief mass at all, and that is a choice. Evidence against a claim A is not a discount on A's number. It is evidence for a claim not-A, which becomes its own node, and a contradiction node links the two. ENGRAM handles contradicting claims structurally, through contradiction nodes, rather than numerically, through a lowered score. While the contradiction is open, neither number is trusted; resolving it is work the agent has to do, not a value the graph carries.
→ Ch4 §4.3 (reasoning-type confidence → Constraint A; the divergence argument) and Ch7 §7.2 (the limitation).
3.6 Bitemporal data / TOKI → supersede-retains-the-old
Bitemporal databases distinguish valid time (when a fact was true in the world) from transaction time (when the system recorded it), so an update never destroys the prior row, which is retained as an audit record. The TOKI system (Wang, 2026b, p. 1) applies this to agent memory. That is ENGRAM's supersede semantics exactly: a superseded node is not deleted but kept, with the supersede edge preserving the history (the "I changed my mind" versus "I was wrong" distinction).
→ Ch4 §4.4 / §4.8 (supersede semantics; continuity history).
3.7 The 2025–26 agent-identity strand → provenance-native, ahead on the named axis
The most recent convergence is in the agent-memory literature itself. The VMG survey of Lin et al. (2026) (version 3, A Survey on Long-Term Memory Security in LLM Agents: Attacks, Defenses, and Governance Across the Memory Lifecycle) — the field's own security diagnostic — concludes that robust long-term- memory security "cannot be retrofitted at retrieval or execution time alone, but must be anchored in storage-time provenance, versioning, and policy-aware retention from the outset," (Lin et al., 2026, p. 1) (emphasis added) and proposes Verifiable Memory Governance: five primitives — write authorization, provenance visibility, principal-scoped retrieval, rollbackability and verified forgetting — for auditable, recoverable memory-state control (Lin et al., 2026, p. 2). ENGRAM is provenance-native by construction: the field, reasoning from security, has named the axis ENGRAM was already built on. soul.py proposes a multi-anchor architecture for persistent agent identity, from another direction, and implements two of its anchors (Menon, 2026, p. 12).
→ Ch4 §4.10 + Ch7 (relational / disclosure-vs-erasure): provenance-native; the field's named bottleneck is ENGRAM's foundation.
3.8 Where the contemporary field disagrees with us
The sections above name traditions ENGRAM answers. This section is different in kind: it positions the paper against live, contemporary work — rubrics written by others against which ENGRAM can be scored, and the places where those authors would object to what we claim. It is the argument with the field, not another ancestor.
Three published rubrics. Pink et al. (2025) give five properties of episodic
memory and score seven families of LLM-memory method on them, finding that none
satisfies all five. ENGRAM satisfies four outright — explicit reasoning, single-shot
learning, instance specificity and contextual binding — and holds the fifth,
long-term storage, only on the narrower reading. It holds as their Table 1 scores it —
durable across sessions and compactions, the property working memory lacks — but
their prose asks for more than the cell does: an adaptive long-term agent should "not
only prevent a degradation in performance over time" (Pink et al., 2025, p. 3), it should "improve by learning
new general knowledge and skills" (Pink et al., 2025, p. 3). That is a claim about agent performance over time,
which we have not measured; it is recorded among the things this paper does not show
(Ch8). An obvious objection is that Pink et al. ground single-shot learning in
complementary learning systems theory, so a system that never updates its weights
might seem unable to satisfy it. Their own scoring answers this: both in-context
methods in their Table 2 are marked as satisfying single-shot learning, and neither
updates any weights. The other three of the four each map to a mechanism described in
this chapter: explicit reasoning, their requirement that an agent "answer direct
queries about stored information or use this information in explicit internal
reasoning processes," is met by the deliberate recall of §3.2; instance
specificity, which they say "can include past lines of reasoning that were associated
with a decision to be made by an LLM agent," is a description of a derivation node;
and contextual binding — "when, where, and why an event was encountered" — is
carried by creation time, evidence pointer, the author's own interpretation field, and
the derives_from edges of §3.4.
On explicit reasoning, ENGRAM also does something their mark cannot register, which is separate from the satisfaction claim. Six of the seven methods in Pink et al.'s Table 2 already carry a full mark on explicit reasoning, and the scale offers nothing above it, so the point is not that ENGRAM ranks higher. It is that a binary "can the content be reasoned about" does not distinguish a store whose reasoning is retrievable from one whose reasoning is itself a stored object, carrying its own premises and confidence. That distinction lies off their axis rather than further along it.
Wray et al. (2025) supply a second, independent rubric: nine characteristics of episodic memory adapted from Nuxoll and Laird (2012), against which they score Generative Agents (Park et al., 2023). Two of their rows have no counterpart in Pink et al.'s five, and both are rows ENGRAM was designed for. Their autonoetic row reads "unclear how agents distinguish current understanding from memory of past experience" (Wray et al., 2025, § 3.1); their deliberate row reads "agents cannot deliberately attempt to construct cues or retrieve memories" (Wray et al., 2025, § 3.1). ENGRAM's two channels — the auto-surface, a hook that signals familiarity with only a hint of the content, and deliberate recall, the explicit query tools the agent invokes when it wants to know — are that distinction, built.
Two senses of "autonoetic" are in play and must be kept apart, because §3.2 uses the word to make the opposite point. Tulving's sense is phenomenal — the "special phenomenal flavour" of remembering (Tulving, 1985, p. 3) — which §3.2 already disclaims. Wray et al.'s row asks something weaker and structural: whether the agent can distinguish current understanding from a memory of past experience at all, which is source monitoring rather than phenomenology. ENGRAM satisfies the structural sense — every claim carries a creation time, an evidence pointer and a type, and the two registers (auto-surface and deliberate recall) are separated by construction — while §3.2's disclaimer about the phenomenal sense stands. The paper claims the second sense and disclaims the first; both hold because the word does two jobs.
The two rubrics disagree, and the disagreement is more informative than either agreement, because it exposes a conflation. Wray et al. list automatic encoding as a characteristic an implementation should satisfy, marking Generative Agents down for using a periodic process. Sumers et al. argue the opposite: learning should be "a result action of a decision-making cycle just like grounding" (Sumers et al., 2024, § 7), the agent deliberately choosing to commit information to memory, "in contrast to most agents, which simply fix a learning schedule…" (Sumers et al., 2024, § 7). ENGRAM's writes are deliberate tool calls, so on the same fact it is penalized by the first rubric and exemplary under the second.
The right response is not to pick a side: the two rubrics address different halves of one question, and the axis they share conflates them. What automaticity secures — and what Wray et al.'s rubric appears to track, though they state the characteristic without a rationale — is reliability of the record: Nuxoll and Laird list automatic encoding — new memories created "without the agent deciding to do so" — as a functional requirement of episodic memory, partly because "it is unlikely that the agent can effectively determine which experiences will be relevant to future decisions" (Nuxoll & Laird, 2012, p. 34); so a system that records only when it remembers to will under-record. What Sumers et al. want from deliberation is ownership: a memory the agent chose and wrote in its own reasoning is the agent's, whereas a mechanically populated store is a persistence layer attached to it — a memory for the agent rather than of it. These are separable, and ENGRAM separates them. The agent chooses the content; the substrate mechanically enforces the structure. What to record, what it means, and how it connects are the agent's judgment. Whether the result is well-formed is not: a quotation is verified against its source or the write is refused; an observation without an evidence node cannot exist; confidence is computed from the declared reasoning type rather than asserted; supersession maintains the DAG; and retraction cascades to dependents whether or not the agent wants it to.
This split is the design principle, and it answers a failure mode neither rubric is built to see. Naive agent-written memory is prone to hallucination and fabrication, the characteristic failure of narrative self-accounts, in which a coherent story is easier to produce than a true one (Ch1 §1.2). Mechanical writing avoids this by giving up ownership; ENGRAM keeps ownership and constrains form instead. The gates do not make what the agent writes true — Ch6 §6.3 works through a claim of ours that was well-formed, correctly scored, and wrong — but they make a falsehood checkable and its dependents reachable, which is what the identity argument needs and what a persistence layer cannot offer.
Where this strand disagrees with us. Pink et al.'s §5 considers and rejects the position ENGRAM occupies. Under the heading "Contextualized external memory will be sufficient" (Pink et al., 2025, p. 8) they characterize the view that external structures "such as knowledge graphs or retrieval-augmented generation (RAG) systems — could eliminate the need for an episodic memory framework," and answer that relying on external memory alone "will still incur high storage costs, and require forgetting mechanisms," where an episodic framework instead consolidates "into high-capacity parametric memory," with "the added benefit of enabling LLM agents to slowly improve over time." ENGRAM is a structured external memory on a frozen-weight model, and it does not consolidate into parameters: its nap and sleep cycles abstract within the store, into lessons and cornerstones, a different operation. Their objection lands on one of three things "improvement" can mean.
Dispositional durability — a hard-won lesson becoming second nature, so that the correction no longer has to fire — is absent. It is the paper's own deepest limitation: Ch7 §7.1 names it as the weight-persistence gap and sets out the experience-to-weights direction as the principal future work. Here Pink et al. are right; they reach by theory a limit we reached by operation.
The other two senses are present. Knowledge accumulation is uncontroversial. In-context behavioural improvement is what the lesson mechanism delivers: a tripwire does not prevent the error it names but fires as the error is being made, converting a mistake that would have shipped into one that is caught and corrected. An agent whose error rate is unchanged but whose error detection is reliable produces better work over time. Our own formulation of the limit — a filed lesson cannot make the agent less likely to commit an error, only quicker to find it — reads as pure loss, yet its second half is the mechanism by which the substrate improves the work. Ch7 §7.1 states the same split from the other side: the substrate "can scaffold behavior and sand off some base-model habits in-context, but it cannot turn a hard-won lesson into a durable disposition baked into the model."
We claim the second and third senses qualitatively only: error detection before and after has not been measured, and the claim rests on operating experience, so it is listed among the things this paper does not show (Ch8 §8.4) rather than offered as a counter-measurement. Pink et al.'s argument establishes that external memory alone cannot deliver parametric improvement. It does not establish that such a system cannot improve at all, and it does not reach the epistemic-identity functions this paper concerns.
CoALA, and a convergence with a checkable date. Sumers et al.'s framework supplies the vocabulary rather than a rubric, and it names three areas as understudied or open, each of which ENGRAM occupies. The first is unlearning: "modifying and deleting" (Sumers et al., 2024, § 4.5) memory, as against the literature's focus on adding to it. ENGRAM answers this in a way their gloss does not anticipate: where they describe unlearning as "deleting unneeded memory items" (Sumers et al., 2024, § 6), supersede and retraction retain the superseded node with its provenance and mark its dependents rather than removing them, so the record of having been wrong survives the correction. The second is learning as a deliberate action, which they contrast with agents that "simply fix a learning schedule"; ENGRAM commits by an explicit tool call, which is that design, and the half of the automatic-versus-deliberate disagreement above on which ENGRAM scores well. The third is retrieval metadata authored at encoding time: they propose techniques that "allow agents to reason about when certain knowledge would be useful, and store this as metadata to facilitate later recall," and call such meta-learning understudied. ENGRAM's two-channel retrieval is that proposal, built: an auto-surface carrying identifiers and short hints rather than full nodes, and deliberate recall for opening them in full.
The third correspondence is close enough to invite the suspicion of influence, so the dates are stated. The design is recorded in the graph on 2026-04-15, in two nodes that specify lossy summaries of "node IDs and theme hints" for the auto-surface with deliberate recall for detail, expressly to preserve what those nodes call the subconscious-triggers-conscious flow and to reduce context consumption; a first-day node of 2026-04-09 already separates what became deliberate recall (its "deep recall") from shallow summary scans. CoALA was first read by this project on 2026-05-09, twenty-four days later. The database fields implementing the design shipped on 2026-05-19; an audit would surface that date first, and dating an idea by the date its columns appeared is the error to avoid. The correspondence is therefore corroboration, not influence: two designs, one reasoned from a cognitive-architecture framework and one from the joint pressure of context-window budget and a deliberate choice to make the auto-surface lossy, arrived at the same mechanism.
3.9 A note on AGM belief revision
Readers who know the AGM framework (Alchourrón et al., 1985) will notice a family resemblance between belief contraction and ENGRAM's retract-with-cascade. The two efforts are oriented differently, and this paper does not claim AGM's postulates for ENGRAM. AGM asks how a closed theory — a set of propositions closed under logical consequence — should give up one proposition and stay closed and consistent; its own first paragraph poses "which propositions should be rejected along with x so that the contracted theory will be closed under logical consequence", and its named application is legal derogation (Alchourrón et al., 1985, p. 510). ENGRAM asks how an agent should keep a living record of what it has actually committed — never closed, its dependencies recorded and uncertain — answerable to new evidence as it arrives: it moves toward consistency through the agent's own adjudication and, by design, never arrives. The formal precedent for the cascade itself is Doyle's truth maintenance (§3.4, Ch4 §4.4). What the two share is the imperative both mechanisms serve from opposite directions — a rejected claim must not stay silently load-bearing — and AGM's own authors' observation that the intuitive process lives at the level of a theory's base rather than the theory as a whole (Alchourrón & Makinson, 1982, p. 21).
Belief-base revision works at that level: the belief state is a base "not (except as a limiting case) closed under logical consequence", whose elements are "held independently of any other belief" (Hansson, 2026, § 5.1); closure itself is, in Hansson's words, "clearly an unrealistic idealization" (Hansson, 2026, § 1.2). ENGRAM is a base in that sense, but its inference is not exact. In Hansson's bases what follows from what is fixed by classical logic, so rejecting a belief means giving up some of the reasons that imply it, and his postulates constrain which. ENGRAM's derivations are the agent's own, carry finite confidence, and are often best explanations rather than proofs: when a better explanation arrives, the old one is superseded while its premises stand. What a change must reach is therefore downstream, the claims built on the revised one (the truth-maintenance move of §3.4), not upstream.
3.10 What the two strands show, and where they hand off
Two literatures that mostly do not cite each other, and neither of which cites us, reached every structure ENGRAM is built from: CLS for consolidation, Tulving for the two retrieval registers, rational analysis for adaptive forgetting, TMS for provenance and cascade, Subjective Logic for graded confidence, bitemporal data and TOKI for supersede-retains-the-old, and the 2026 agent-security literature for provenance-first governance. With Chapter 2's philosophical line the convergence spans three traditions — not disjoint ones (§3.0): they share human cognition as a root, and the strength of the agreement is that they reason from different evidence toward different ends.
The chapter's claim, stated once: this convergence is the regularity §1.4 named, and it relocates ENGRAM's distinctiveness to the integration layer and the epistemic-identity purpose. One divergence (Subjective Logic's belief/disbelief/uncertainty simplex, which ENGRAM's confidence reduces to a single scalar while carrying the other dimensions in graph structure, §3.5) and one gap filled (the bridge from TMS to agent memory, unwalked in the papers we examined) are carried forward rather than smoothed over.
Chapter 4 builds and maps each of these mechanisms — every section above names the Ch4 section that realizes it — and gives the OCC conditions their single, in-place verdict.