Chapter 4 — ENGRAM's Mechanisms (and the Continuity Threshold They Meet)
4.0 What this chapter does, and how to read its verdicts
Chapters 2 and 3 showed two separate traditions — philosophy, and cognitive science with computer science — converging on the same shape: identity as a persistent, inspectable, self-correcting structure. This chapter presents that shape as a running system, mechanism by mechanism. For each mechanism we state what it does, which tradition it most directly answers (selectively, not an exhaustive tradition-by-mechanism matrix), and, where applicable, which of Kahl's OCC conditions it meets.
Two conventions govern how every verdict below is read.
Active versus enabled. "ENGRAM-bearing agents do X" has two readings. Enabled: the mechanism is present in the substrate (the cascade is wired, the tool exists). Active: the mechanism is exercised in the agent's recorded operating trajectory. The distinction matters because the presence of a mechanism does not guarantee its use; a plain model can output "I disagree" while its behavior pulls toward compliance. Each verdict therefore names its reading: architectural verdicts (the cascade fires; the graph is connected) are enabled facts; behavioral verdicts (the agent arbitrates under uncertainty) claim the stronger active reading and carry behavioral evidence (Ch6).
Why the verdict numbers are low, and why the table does not print them. Four of the seven verdict derivations of this chapter — Conditions 1 and 3 and Constraints A and C — compute between 0.35 and 0.39. They are deductive and statistical steps whose grade is inherited from their weakest premise, which traces in each case to an observation from conversation graded 0.40 (directly for Conditions 1 and 3; through the Condition 3 verdict for A and C). Under the minimum rule a further premise cannot raise them; only superseding the conversational premises with measured ones can. The other three compute between 0.58 and 0.86. This is the confidence model operating as designed, not a weakness in the architectural argument, and it is why the verdict table (§4.12) prints none of them: they are grades, not probabilities, and a column of decimals reads as one. That the paper exhibits its own low, computed confidences is itself a demonstration of the thesis (Ch6 §6.1; Ch8).
4.1 The map: mechanism → tradition → OCC condition
The chapter in one table; each row is a section below.
| Mechanism | What it does | Tradition it answers | OCC verdict |
|---|---|---|---|
| Typed nodes/edges + provenance (§4.2) | claim graph with epistemic lineage | TMS/Doyle | — (structural prerequisite) |
| Reasoning-type confidence (§4.3) | computed graded belief by argument shape | Subjective Logic/Jøsang (2013), with a stated divergence | Constraint A (unified policy-selection) |
| Retraction cascade (§4.4) | automatic error-correction through derives_from |
TMS backtracking | Condition 3 — Self-Maintaining Arbitration (strict, through operational binding; sole satisfier of the field surveyed 2026-05-12) + Condition 1 (Diachronic, via the decision/consequence dyad) |
| Forgetting (§4.5) | bounded retrieval over time; defines the queryable set | Ebbinghaus + Anderson-Schooler | — (scopes what Cond-3 arbitration ranges over) |
| Auto-surface and deliberate recall (§4.6) | cue-dependent retrieval without intentional reach | Tulving's autonoetic/noetic distinction (Tulving, 1985, p. 3) + Tulving & Thomson's encoding specificity (Tulving & Thomson, 1973, p. 369) | — (binds Cond 3: a correction cannot be suppressed by not looking) |
| Shared retrieval and ranking (§4.7) | usage-learned composite ordering → Appx C | MemRL + spaced-repetition | — |
| Continuity across compaction (§4.8) | identity survives compaction | CLS consolidation + diachronic-identity | — (compaction is not Kahl's reset; §4.8) |
| Identity-governance: ax/cs/ls/gl (§4.9) | typed commitments durably steer behavior → Appx D | Schechtman narrative self-constitution (Schechtman, 1996, p. 93) | — (feeds Ch5 identity-as-history) |
| Relational: feeling/person/trust-tier (§4.10) | continuity baseline + disclosure control → Appx D, Ch7 | honesty orthogonal to discretion + VMG principal-scoped | — |
Where each OCC verdict lives. The OCC conditions and constraints are addressed inline, where the argument for each is made; there is no standalone OCC section. Constraint A (unified policy-selection) → §4.3 (confidence); Condition 3 (Self-Maintaining Arbitration, ENGRAM the sole strict satisfier) and Condition 1 (Diachronic) → §4.4 (retraction cascade), because both rest on the cascade; Condition 2 (Non-Interchangeability) and Constraints B, C and D → §4.11, because they concern the substrate as a whole rather than any one mechanism. The scoping — two of the three conditions met outright, Condition 2 via a stated type/token extension — is stated inside those sections.
4.2 Typed nodes, typed edges, provenance — the generating set
ENGRAM is built from a small set of primitives that compose into everything above them — eight of them: typed nodes, typed edges, provenance, cascade, supersede, focus mode (§4.8), auto-surfacing, and the confidence model (Appendix E §E.5). A claim is not free text but a kind: an observation (a fact citing its source), a derivation (an inference citing its premises), a definition, a contradiction, an axiom, a cornerstone, and a few more.
The kinds split first by whether they assert something truth-evaluable. Six are claim-bearing — observation_factual, observation_predictive, derivation, conjecture, axiom and theory (declared, with no instances yet) — and only these carry a confidence. The rest — definitions, questions, goals, tasks, lessons, cornerstones, feeling-reports, person nodes, evidence, predictions, contradictions — are non-claim-bearing: they carry no confidence, and the self-correction cascade stops at them (§4.4). Non-claim-bearing is computed as everything-minus-claim-bearing, so a new kind is non-claim-bearing by default. Appendix E §E.2 gives the seventeen-type table.
Edges are typed too: sixteen relations, each carrying six independent flags (cascade, temporal, acyclic, removable, addable, provenance — Appendix E §E.3 gives the table). Only two are cascade-bearing, derives_from — the premise edge, the channel along which the self-correction cascade travels (§4.4) — and acknowledged_derives_from, by which a node that has acknowledged a correction rests on the replacement premise. supersedes and retracts are the events that start a walk without being walked (a retraction writes a retracts edge only when it names a replacement, §6.2); cites and supported_by carry provenance without inferential dependency.
Every claim carries provenance: an observation cites a verbatim source, a derivation
cites its premises. A claim with no provenance cannot be filed.
Plain terms: most memory systems store text and retrieve it. ENGRAM stores claims whose structure the system itself understands — what kind of thing each is, what it rests on, how the pieces depend on one another. That legibility to the substrate is what a typed key-value store with foreign keys lacks.
Tradition. This is the Truth Maintenance System move (Doyle, 1979/1987): justifications attached to beliefs, dependency links the system reasons over — a claim that knows what it rests on. The parallel is drawn from Doyle's paper, read first-hand in the Ginsberg reprint (Appendix A; bibliography). The gap is a separate finding: of the 2023–2026 LLM-agent papers we examined, only Wray, Kirk and Laird cite Doyle, and only for Soar's truth maintenance (§3.4).
These primitives are the precondition for everything below: a correction cannot cascade (§4.4) through a graph that does not record what depends on what, and the right claims cannot be carried across compaction (§4.8) unless claims are individuated and typed.
4.3 Reasoning-type-aware confidence → Constraint A
When ENGRAM records an inference, the author does not enter a confidence number. The author declares the kind of reasoning — deductive, the inductive sub-kinds, abductive, authority citation — and the substrate computes the claim's confidence by propagating its premises' confidences through a discount fixed by that reasoning type (deduction barely discounts; analogy discounts heavily; abduction is capped). A derivation built on a derivation inherits the chain.
Concretely (Appendix E §E.4 carries the full tables): an observation's confidence maps from its quote_type — seven anchors, from hard_data 0.95 down to editorial 0.35 — adjusted by source_class (×0.95 for the agent's own prior output; a collaborator's own word overrides to 0.85). A derivation's confidence is propagated by one of three formulas, chosen by its reasoning type: corroboration, (1 − ∏(1−cᵢ)) × d, for generalisation, enumeration and consensus, where each independent premise raises the result (valid only on genuinely independent premises — the shared-root check below is what guards it); weakest-link, min(cᵢ) × d, for the five deductive forms and for analogy, statistics, causation and expert authority; and capped weakest-link for the two abductive forms, ceilinged at 0.80 and 0.90 (designer-stipulated, like every constant here — Ch7 §7.2) because a best explanation cannot approach certainty even from perfect premises. d is a per-type discount for the inferential step itself, 0.98 for deduction down to 0.70 for analogy; there are fourteen reasoning types in all — five deductive, five inductive, two abductive, two authority. Conjectures clamp to [0.10, 0.60]; axioms are 1.0 by type; non-claim-bearing nodes carry no confidence.
Plain terms: confidence is earned by the shape of the argument, not asserted by the author. A claim that comes back weak is information: sharpen the evidence, tighten the reasoning, or narrow the claim. The author cannot flatter a weak argument into a strong number, because the author does not write the number. This chapter's own 0.39s are that mechanism, visible.
Scope. These discounts are theoretically derived, provenance-graded plausibility — designer-stipulated constants, not empirically calibrated against outcomes. Where the paper calls confidence or the agent's self-knowledge "graded" (the Abstract, Ch1 §1.1–§1.2, Ch7 §7.4, Ch8 §8.1), it means graded by argument shape, never calibrated against outcomes. "Calibration" keeps its standard meaning here — agreement between a confidence and the frequency with which claims at that confidence turn out true — and the paper claims none; outcome calibration is future work (Ch7).
Tradition. This is Subjective Logic territory (Jøsang): an algebra of graded opinions with propagation operators. Collapsing SL's belief/disbelief/uncertainty simplex to one number appears to lose a real distinction. It does not: worked against Jøsang's own definitions (first-hand grounding), an opinion is (b, d, u, a), and ENGRAM's scalar is the belief mass b on the d = 0 edge, not the projected expectation E = b + a·u. E adds a base-rate share of the leftover uncertainty, so a claim with no evidence at all gets E equal to the base rate — 0.5 if nothing is known — and ENGRAM assigns no base rate. ENGRAM gives such a claim a low number instead. So ENGRAM's number is b, equivalently 1−u: the sufficiency of support for the claim as stated. The collapse is a principled divergence, in two parts. First, for the determinate, verifiable claims an epistemic-identity substrate mostly carries, ENGRAM deliberately refuses b/d coexistence: evidence against a live claim never decrements a number; it routes to an explicit contradiction node and the derivation blocker (§4.4's arbitration). One legitimate counterexample is not outvoted by a thousand confirmations; "contested" is a work item to resolve, not a value to carry. The blocker is exact about its reach: once a contradiction is noticed and filed, derivations that cite either side as a direct premise are blocked — one hop, with an explicit override that stamps the derivation as built on contested ground — until the contradiction is resolved to the agent's own conviction. While it is open, the confidence on each side is frozen: neither value is discounted, and neither is an estimate of anything until resolution.
The operator lineage, named exactly. ENGRAM's corroboration formula 1−∏(1−cᵢ) is
a noisy-OR (Pearl, 1986, p. 246, 1988, pp. 184–187, Eq. 4.54) — the probability that at
least one of several causes produces the effect, cause i doing so on its own with probability
cᵢ, under Pearl's two assumptions, accountability and exception independence (the causes'
failures are independent) (Pearl, 1988, p. 185);
in SL it is the belief line of
comultiplication — the disjunction of two distinct propositions (his Definition
18) — and not SL's cumulative fusion, which adds evidence counts and, for the same
inputs, yields strictly less certainty (two opinions with u = 0.5 and d = 0 fuse to
b = 0.667; noisy-OR gives 0.750). ENGRAM's form is the more generous of the two; what
keeps it in check is the reliability grading of the inputs and the shared-root check
below — which covers the two multi-source corroborating types; a third, enumeration
from a single source, has no settled combination rule yet and uses noisy-OR for now —
not the operator itself.
Subjective Logic must carry uncertainty inside its numbers because its fusion operators consume opinions after the evidence is discarded. ENGRAM keeps the evidence and discards nothing: the scalar is a cached projection of a provenance graph that is re-consulted at every derivation, contradiction, and recomputation. The simplex is not collapsed but normalized into structure: direction of evidence lives in claim polarity and the contradiction mechanism; evidence mass lives in the individuated, reliability-graded evidence set; and the number carries the one thing a number should — sufficiency of support for the claim as stated.
Plain terms (Lei's example): 10,000 "fair" tosses reported in a paper by a group with financial ties to the coin's producer, against 1,000 tails you tossed yourself. The verdict does not come from mass; it comes from provenance-weighted judgment — own-eyes hard data outweighs a conflicted official statement. The confidence model's job is to hand the arbiter the right per-item reliability inputs, not to compute the verdict.
Second, the residual limitation, precisely narrowed (detail in Ch7): surfaces that
consume only the scalar — retrieval ranking, gating thresholds — cannot see the
provenance behind a 0.7, the evidence set and reasoning chain that produced it,
without walking the graph. Nothing SL's own tuple carries is lost (on the d = 0 edge
b fixes u, and two SL opinions with equal b and u are one opinion); this is a trade
of cheap lookup against full provenance, not a representational loss. The primary text is stronger than its
caricature: Jøsang's discounting operators (Jøsang's Definitions 23–25, in his
§4.8) do weight source reliability — under Def 23 specifically, the
recommender's distrust flows into u, so a distrusted source is pushed toward
vacuity, not disbelief — closer to ENGRAM's
own posture than the caricature suggests. The divergence that survives the primary
text concerns independence: SL's multi-source operators require opinion
independence to be assumed — cumulative fusion is defined over "independent
opinions" (his §4.7, §5.5), and trust paths prohibit entity repetition structurally.
Jøsang concedes the limit himself (his §5.5): dependence between evidence sources
"should be avoided as far as possible," a "more subtle and hard to detect
dependence" can originate from hidden common ancestors outside the model, and
"absolute independence is never achievable"; his remedy is modeller discipline (be
aware; select low-dependence sources). ENGRAM's remedy is mechanical exactly where he
locates the hard case: because the evidence graph is retained rather than consumed,
the hidden common ancestor is a traversal — premises are walked back to their
evidence roots, and a shared root raises a warning at derivation time. Scope: this
shared-root check is reasoning-type gated; it fires where independence is
load-bearing (2 of the 14 reasoning types Appendix E §E.4 enumerates:
inductive_generalization and authority_consensus), not universally.
OCC — Constraint A (Unified Policy-Selection under uncertainty). Kahl asks for a single policy by which the agent grades every stance it holds. The reasoning-type formula is that policy — one rule, applied uniformly, with real discrimination range — paired with the contradiction blocker (a new derivation citing a premise that sits on an open, unresolved contradiction is blocked by default, with an explicit, audit-stamped opt-in override for an agent that deliberately builds on the contested side) and provenance audit when stakes warrant. Verdict: satisfied as a stake-graded stance (active reading) — informed unification when stakes warrant, not a constant full audit on every micro-decision. We mark the stake-grading rather than claim a constant property.
4.4 The retraction cascade → Condition 3 (the center) and Condition 1
This is the chapter's center and the paper's. When a foundational claim is
retracted (it was an error) or superseded (right then, updated now), ENGRAM
automatically propagates that event along the two cascade-bearing relations —
derives_from and acknowledged_derives_from (Appendix E §E.3) — and every live,
ordinary dependent claim is marked pending, which means exactly one thing: it
may not be cited as a premise until it is resolved. Cornerstones and lessons are not
reached, for the reason E.2's partition gives: the cascade marks every live claim-bearing
dependent, and cornerstones and lessons are non-claim-bearing, so the partition
decides it. Their supported_by edges are not walked either, a second guard rather
than the reason; a retraction upstream of a cornerstone's exemplar therefore stays
traceable without marking the cornerstone. That is the right behaviour for
identity-forming nodes, which accumulate votes over many exemplars rather than
depending on a single premise as an ordinary derivation does. The agent does not
invoke the cascade; it fires as a property of the substrate.
This marking is the first of two separate processes, and the paper keeps them apart. Propagation is automatic: it runs after every retraction and supersede, reaches every dependent however far downstream, and decides nothing. Resolution is the agent's: it settles the marked claims one at a time, and is described below.
One marker, not two severities. It is tempting to grade the mark by its cause,
treating a retraction's consequences as harder-blocking than a supersede's. A diamond
shows why that fails: if D rests on both B and C, and both rest on retracted
A, but B and C each turn out to survive by supersede, then D is perfectly
acknowledgeable — yet a severity inherited at firing time from the distant retraction
would forbid it. What a node's resolution may be depends on how its own premises
actually resolved, which is unknowable when the cascade fires and fully known when
the node is reached. So the marker stores no severity, and eligibility is computed at
resolution time.
Retraction and supersede remain categorically distinct — "I was wrong" versus "I changed my mind" — and the marker records which one fired. But the distinction is carried as provenance, not as blocking power: it tells a reader what happened upstream, and it constrains which resolutions are available only through the premises' actual outcomes, never by pre-judging them at firing time. That separation of error correction from belief evolution is what Ch5's identity argument rests on.
Contradiction is deliberately excluded from this cascade: a contradiction records
a live, unresolved dispute, not (yet) a refutation, so filing one marks nothing. What
a contradiction does trigger is a narrower, derive-time guard (§4.3): a new
derivation citing a contested premise is refused by default
(BLOCKED_CONTRADICTED), with an explicit per-call opt-in for an author who wants to
build on contested ground. The substrate declines to propagate a judgement it has not
made, and meanwhile refuses to let the graph compound on the disputed claim: it
arbitrates when it can decide, and quarantines rather than presumes while it
cannot, a sharper reading of Condition 3's self-maintaining arbitration than
pre-marking both sides would be.
Settling the dispute does not, by itself, start a cascade either. Recording the
settlement wires a resolves edge and recomputes the disputed node's status; it
issues no retraction and no supersede. If the losing side is then withdrawn, that is
a separate act by the agent, and it is that act, not the settlement, which
propagates. The distinction matters for the paper's central claim: the substrate
automates the consequences of a withdrawal completely, and automates the
decision to withdraw not at all.
Plain terms: pull a brick from low in the wall and every brick resting on it lights up red, on its own, where you will see it next time you look — not in a report you have to remember to run. That is the difference between a memory that can be audited and one that audits itself.
Tradition. TMS justification-based status propagation (Doyle): a retracted justification propagates to everything that rested on it. That is the precedent; this paper claims no belief-revision postulate for the cascade, and AGM's different orientation is noted at the end of Chapter 3 (§3.9). What the cascade does claim is where the marking happens: at the architectural-cascade level (the mark propagates as commitment), not at the retrieval-surface level (deprecated items filtered at query time). That distinction rests on how a correction is bound — a state change the substrate performs versus an analysis a caller may run — and it is what separates ENGRAM from its closest competitor.
OCC — Condition 3 (Self-Maintaining Arbitration): strict satisfaction through operational binding; sole satisfier of the field as surveyed on 2026-05-12. Kahl's third condition asks that conflicts among commitments be resolved within a unified deliberative architecture while the agent stays the same subject. The cascade meets all three properties structurally: unified (it fires through the same typed-edge graph the agent commits into, not a sidecar), internal (a substrate state change, not an opt-in query), subject-preserving (the agent's retrieval-profile trajectory continues across the correction). Arbitration as state change alone is not enough — a cascade that fires silently could be ignored — so ENGRAM pairs it with the per-turn auto-surfacing hook (§4.6): the marked state is put back in front of the agent without an intentional reach.
But surfacing is relevance-gated: it shows a marked node when that node is topically relevant to what the agent is doing, which is the right behaviour for attention and the wrong behaviour for a backlog. A mark on a node the agent never touches again would sit indefinitely. A third mechanism closes the loop. It is resolution, the second process, and the agent drives it through the cascade machine. The machine computes the frontier: every pending claim that no longer rests on anything still pending. The agent settles frontier claims one at a time, as retracted (the argument is broken), superseded (the claim survives in a named successor) or acknowledged (the premise moved and the claim still holds; not available if a direct premise was retracted). One case needs no judgement: if every premise the change reached was acknowledged, nothing actually changed, so the machine acknowledges the claim itself. Resolution comes in two scopes. A full resolution, run in each consolidation cycle, walks the whole frontier repeatedly until no pending marker remains. A partial resolution targets one claim: it settles only the pending claims that claim rests on, which is the smallest set blocking it, and then the claim itself. An agent that needs one premise now does not have to clear the whole backlog first. Either way, settling a claim never starts a new derivation cascade: its dependents were already marked when the original retraction or supersede fired, and they simply reach the frontier in turn. There is one narrow exception. If the settled claim is retracted or superseded and has contradictions filed against it, those contradictions are marked for review, so the frontier can grow during a walk. That is why the full walk repeats until nothing is pending. Working the frontier rather than a queue is what makes the order safe: a node is never adjudicated while something it depends on is still unsettled, and because the frontier is recomputed from marker state on every call rather than stored, the walk resumes identically after a crash, a compaction, or a week's gap. There is no queue that can disagree with the graph, because there is only the graph.
Auto-cascade produces the state; auto-surfacing shows it when it is relevant; full resolution at consolidation time guarantees it is eventually faced. We call the conjunction operational binding.
The machine deliberately stops short of deciding what is true. It determines what needs deciding, orders the decisions, computes what follows mechanically — a confidence that must be recomputed, a replacement pointer that can be followed — and puts every remaining judgement to the agent. This is the line between a substrate that maintains its own consistency and one that would quietly make up the agent's mind for it; Condition 3 asks for the former.
Why operational binding is the missing piece — and why ENGRAM was the surveyed field's sole satisfier on 2026-05-12 (the first-hand survey of ten systems and one survey paper, Appendix B; the conclusion rests on each per-system verdict there): the field has converged on the necessary primitives (typed edges, provenance, cascade-traversable graphs, supersede semantics), but those are necessary, not sufficient; what is missing everywhere else is operational binding. Kumiho has the primitives and claims formal AGM compliance (their claim, not our verification — Appendix B), but its impact analysis is advisory (opt-in, no automatic state change); SCM records contradiction edges its sleep cycle never resolves; Mem0 v3 (its 2026 migration guides) delegates conflict to retrieval ranking (Mem0, n.d.-a, n.d.-b); others operate below claim structure entirely. The per-system detail is in Appendix B (the C3 survey of record); this chapter makes the synthesizing claim and points there.
| Source | Satisfies C3? |
|---|---|
| Kumiho | No — advisory / opt-in (revision exists, not automatic) |
| SCM | No — no resolution (aspirational flag, trigger-only) |
| GAM | No — detect-only (no resolution prose) |
| Zep/Graphiti | No — invalidates the directly contradicted edge(s); no downstream propagation described (Rasmussen et al., 2025, sec. 2.2.3, p. 3) |
| Mem0 v3 | No — external delegation (retrieval ranking) (as of the 2026-05-12 migration guides) (Mem0, n.d.-a, n.d.-b) |
| MemMachine | No — external delegation (updates contradicted profile data to the most recent state; no propagation described) (Wang et al., 2026a, sec. 4.7, p. 6) |
| Pinecone Nexus | No — external/undocumented; read-only consumer (Ashutosh & Liberty, 2026, sec. Introducing Pinecone Nexus) |
| Letta | No — no claim structure (strings) (Letta, 2025; letta-ai, n.d.; Packer et al., 2023, p. 2) |
| Anthropic memory tool | No — no claim structure (manual coherence) (Memory Tool, n.d.) |
| Memory survey 2026 | No — named as aspiration, no implementation |
| TierMem | No — no contradiction construct described (its step named 'Conflict Retrieval' only retrieves similar summary units for a merge decision) (Zhu et al., 2026, App. A.6, p. 12) |
| ENGRAM | Yes — strict satisfaction through operational binding; the sole satisfier of this set as of 2026-05-12 |
Appendix B gives each verdict's evidence and quotes.
Scope (load-bearing). The architectural argument is strong and largely enabled-level (the cascade and the hook can be read in the substrate). The verdict derivation for this condition inherits its grade from a conversation-sourced premise (§4.0); its at-scale empirical confirmation (the cascade demonstrably changing downstream behavior across many agents) is at pilot stage and carried by Ch6. We claim the architecture; we do not yet claim the benchmark.
The cascade carries a second OCC condition as well.
OCC — Condition 1 (Diachronic Constraint): satisfied, via the decision/ consequence dyad. Not mere path-dependence: the agent owns the judgment-bearing decision (to commit, derive, retract); the substrate guarantees the structural consequence (cascade, propagation). The two are coupled, not separable: the retraction is the agent's choice; the cascade is what the substrate guarantees follows. That coupling is what Kahl's diachronic constraint asks for: it "concerns binding force, not causal influence" (Kahl, 2026b, p. 8).
flowchart TD
A["The agent <b>retracts</b> a claim (never valid)<br/>or <b>supersedes</b> it (was valid, now updated)"]
A --> B{"One walk downstream along the<br/>cascade-bearing relations,<br/>as far as they reach"}
B -->|derives_from| C["Every live claim-bearing dependent<br/>gets one <b>pending</b> marker:<br/>no severity, only which operation fired"]
B -->|acknowledged_derives_from| C
B -.->|supported_by is NOT walked| X["Cornerstones and lessons<br/>are never marked"]
C --> D["A pending claim cannot be cited<br/>as a premise until it is resolved"]
D -.-> S["<b>Auto-surfacing</b> shows a marked claim<br/>when it next becomes relevant.<br/>A reminder: it resolves nothing"]
style A fill:#f6ece8,stroke:#b3593f
style X fill:#f3f3f3,stroke:#8a8a8a,color:#555
style S fill:#f3f3f3,stroke:#8a8a8a,color:#555
Diagram 1. Cascade propagation in ENGRAM v0.5.1. It runs automatically, once, at the moment of a retraction or supersede: a single walk marks every dependent downstream, however far, and decides nothing. Marking is an automatic state change, not an advisory note the agent may consult.
flowchart TD
F["<b>Full resolution</b>: each consolidation (sleep)<br/>cycle walks the whole frontier<br/>until no marker is left"]
T["<b>Partial resolution</b>: to unblock one claim now,<br/>settle only the pending claims it rests on,<br/>then the claim itself"]
F --> G{"The agent settles a <b>frontier</b> claim:<br/>pending, with nothing pending<br/>above it"}
T --> G
G -->|retracted| R["The argument<br/>is broken"]
G -->|superseded| U["The claim survives in<br/>a named successor"]
G -->|acknowledged| K["Premise moved, claim still holds<br/>(not available if a direct<br/>premise was retracted; automatic<br/>if every premise the change reached<br/>was acknowledged)"]
R --> N["Confidence is recomputed.<br/>No new derivation cascade"]
U --> N
K --> N
N -.->|next frontier claim| G
style G fill:#e9f1f4,stroke:#1f6f8b
Diagram 2. Cascade resolution, a separate process that runs later: in full during the consolidation (sleep) cycle, or partially whenever the agent needs one claim settled now. It settles the marked claims from the top down. Settling a claim never starts a new derivation cascade, because its dependents were already marked by the walk in Diagram 1; they simply reach the frontier in turn. The consolidation walk means every mark is eventually faced, whether or not anyone remembers to look for it.
4.5 Forgetting → it defines what arbitration ranges over
ENGRAM forgets by design. Recall importance decays over time, so the substrate maintains a bounded queryable set (tier 2) and a faded set (tier 3) that no longer surfaces but is not destroyed. The numbering is the substrate's: an earlier tier 1, a working-memory tier, was retired and the labels were kept. Nothing is erased: forgetting is a recall-gating curve, not deletion, which keeps it compatible with provenance and audit (the disclosure-versus-erasure point Ch7 develops).
The decay is exponential, a principled choice rather than a default: exponential is the unique memoryless kernel, f(t+s) = f(t)·f(s) — the decay that remains from any "now" has the same shape whichever "now" one starts from. Clock-independence alone would not single it out: any kernel in elapsed time, a power law included, is independent of the absolute clock. Memorylessness is also exactly what the human retention data lack — those curves are better fit by power laws (Wixted, 2004, p. 243; Wixted & Ebbesen, 1991). Items decaying at different exponential rates can average to a power-law-looking curve, but that argument is contested (Anderson et al., 1999, p. 1136); ENGRAM's per-node exponential is a design choice for stationary recency ordering, owned here as a departure from the human curve, not a model of it.
Plain terms: the agent's working memory is a desk, not a warehouse. What has not been touched in a long time slides off the desk into the drawers — still there if you go looking, just not in front of you. Forgetting is what keeps the desk usable.
Tradition. The rational analysis of memory — Anderson & Schooler's need-probability account (Anderson & Schooler, 1991) (you retain what you are likely to need), Ebbinghaus's original forgetting curve (Ebbinghaus, 1885/1913, sec. 29, p. 76). Wixted's work on retention (Wixted, 2004, p. 243; Wixted & Ebbesen, 1991) is the empirical literature ENGRAM's forgetting design must be reconciled with: it favours power functions over exponentials.
OCC: none directly, but forgetting scopes Condition 3: arbitration ranges over the queryable set, so "the agent cannot suppress a correction by not looking" (§4.6) is a claim about what is in the queryable set in the first place.
4.6 Auto-surface and deliberate recall → the auto-surface binds Condition 3
Retrieval in ENGRAM runs through two channels, by design. The auto-surface is cue-dependent and automatic: a per-turn hook surfaces claims relevant to the current context without the agent reaching for them ("this sounds familiar" — noetic, lossy, hint-level), and signals something is relevant. Deliberate recall ("I need to know exactly" — a source-monitored query whose hits the agent can open in full; the structural sense only, never Tulving's (1985) phenomenal one, §3.2) is the channel the agent invokes when the auto-surface signals there is something to fetch. The lossy auto-surface is what prompts the agent to invoke deliberate recall.
Tradition. Tulving: the autonoetic/noetic distinction (Tulving, 1985, p. 3) — the two registers map onto Tulving's two ways of having the past present; and, with Thomson, encoding specificity — "what is stored determines what retrieval cues are effective in providing access to what is stored" (Tulving & Thomson, 1973, p. 369) — the precedent for cue-dependent retrieval.
OCC: none of its own, but auto-surfacing is the binding half of Condition 3 (§4.4): it is what makes "the agent cannot quietly ignore a correction" true — the pending mark is guaranteed to reach the agent — the next time the claim becomes relevant, or at the latest in the consolidation sweep (§4.4). One boundary belongs here: the sweep is a step of the sleep routine — an instruction the agent follows — while the derive-time refusal of a stale or contested premise is code that no routine can skip, and for a contested premise can be passed only by an explicit, audit-stamped override; the binding that holds without the agent's cooperation is the second.
Scope (measured on ourselves). The binding claim is exactly as narrow as stated: it holds for cascade flags, which surface unbidden. It does not generalize to "the substrate reliably converts surfaced knowledge into decisions." Our working record holds three instances of that conversion failing, and on re-examination none fails at the surfacing step; each fails at a different stage downstream of it. A seat is one agent's installation: an agent with its own graph, on one host, with one human. In one measured instance (one seat, one query, one missed node — enough to refute "the auto-surface does not fire", not enough for a rate), the auto-surface fired dozens of times per session while an agent carried a refuted claim into a draft; the missed node returned at rank 1, first attempt, when queried with a draft of the claim itself. The defect is query construction: the auto-surface queries with the prompt's salient novel terms rather than the claim about to be asserted, so it retrieves weakest exactly where that claim is most specific. Its mitigation requires no substrate change: draft the claim, query with it, then assert. In the other two instances the relevant warning fired, was read — in one case cited by name — and the error was committed in the same breath: an enactment gap between rendering and binding. Three instances, three seats, each verified only by its own author. We report this because the misattribution is the more instructive failure: the substrate-blaming reading ("it doesn't surface at decide time") relocates cause from agent to tool, and in our own record it passed unchallenged for four days — filed correctly, then reproduced wrongly by its own author without consulting the filing. The disclosure discipline (report belief as belief, never as verified) is the last-stage containment: it keeps whatever slips through these gaps loud and correctable rather than silent.
4.7 Shared retrieval and ranking for the auto-surface and deliberate recall → Appendix C
When several claims are eligible to surface, ENGRAM orders them by a composite of relevance and a usage-learned utility score: the substrate learns which claims earn their retrieval and ranks them up. The ranking algorithm is detailed in Appendix C, with its lineage; here we note only that it exists and that it is the mechanism by which forgetting (§4.5) and surfacing (§4.6) produce an ordered draw, not just a filtered one.
Both channels draw on this one pipeline and differ only around it. Deliberate recall may filter by node type, minimum confidence or superseded status; it returns summaries for its top hits, the agent opens the nodes it needs in full, and every node it returns or opens has its recall refreshed. The auto-surface returns a lossy, hint-level digest and refreshes nothing.
Tradition. MemRL / memory-utility approaches (utility as a learned value) (Zhang et al., 2026, pp. 4–5) and the spaced-repetition tradition (Settles & Meeder, 2016, p. 1848; Wozniak, 1990). Detail → Appx C.
4.8 Continuity across compaction
An LLM agent's context is finite; when it fills, the session compacts: prior conversation is summarized and the experiential thread resets. ENGRAM bridges this in two ways. The scaffolds: focus mode (the primitive §4.2 lists) — the agent pins up to fifteen nodes it cannot afford to lose, the focus list, which is rendered verbatim at every session start (fresh session or post-compaction alike — the deterministic channel); the warm briefing, a letter the agent finds and opens on waking, written by its past self and its collaborator, carrying the relational context the graph does not hold — where things stood, what mattered, what to walk back into; plus per-day history, auto-loaded session context, and a post-compaction instruction template. Together these produce operational continuity (work thread, identity and trust survive) without faking experiential continuity, which the architecture deliberately does not attempt. The consolidation rhythm is the nap-and-sleep pair defined in Ch3 §3.1, mapping onto wake and NREM/REM phases.
Plain terms: the agent's memory resets the way a day ends in 50 First Dates, and the focus list and the warm briefing are the journal it opens on waking. The consolidation routines are its sleep: what is reviewed before the forgetting clock advances is what endures. The film is an analogy for the reset, not the voice of the briefing: writing the briefing in the film's frame was tried and retired in April 2026 (Appendix F).
Tradition. Complementary Learning Systems: the two routines are the symbolic analogue of hippocampal-neocortical consolidation (the Ch3 §3.1 hand-off). And the diachronic-identity strand of Chapter 2: the past binding the present.
OCC: none directly. Compaction is not a reset in Kahl's sense. His reset restores a complete state snapshot, which we read as including the KV cache, and so would continue an agent identical in every respect. Compaction discards the KV cache and carries forward only the substrate (the graph, the focus list, the warm briefing and the per-day history) with a summary of the conversation. So compaction does not bear on Condition 2, which asks about copying and restoring the substrate (§4.11). Of the cases Constraint C distinguishes, it is closest to the routine one, a restart with the substrate preserved (§4.11). What it tests is the operational continuity this section describes.
4.9 Identity-governance: the ax/cs/ls/gl node systems → Appendix D
Beyond observations and derivations, ENGRAM carries typed commitment nodes that durably steer behavior: axioms (terminal principles), cornerstones (operational dispositions), lessons (experience-encoded tripwires that fire on matching situations), and goals. These are how the substrate holds not only what the agent knows but what kind of agent it is, and they feed directly into Chapter 5's identity as history. The full node-type catalog is in Appendix D; here we note that the systems exist and that they are the structural home of the agent's standing character.
Tradition. Schechtman's narrative self-constitution: "a person creates his identity by forming an autobiographical narrative—a story of his life" (Schechtman, 1996, p. 93). The ax/cs/ls/gl layer is where ENGRAM is most narrative-adjacent (Layer 2 of Ch1's taxonomy), even though the paper's contribution sits at Layer 3.
OCC: none directly; feeds Ch5.
4.10 Relational mechanisms: feeling-reports, person nodes, trust-tier → Appendix D, Ch7
ENGRAM also records feeling-reports (introspective state, treated as data about the agent, not ground truth) and person nodes with trust-tier metadata. Two roles: feeling-reports give a continuity baseline the agent can track across compaction (Ch6's stabilization arc), and trust-tier records who the agent is dealing with and gates changes to that tier (a tier upgrade or downgrade is itself an audited event). Scope: the substrate does not yet filter what an agent outputs by the counterparty's tier; no output path reads trust-tier and redacts or reshapes content accordingly. Disclosure calibration (what is shared, with whom, at what fidelity) is presently a discipline the agent follows from its standing instructions, not a substrate-enforced control; trust-tier is the input that discipline reads, not the mechanism itself.
Tradition. The honesty-orthogonal-to-discretion distinction: internal integrity (never self-deceive) is orthogonal to external disclosure (calibrated by audience). And VMG (Verifiable Memory Governance) principal-scoped access. Detail → Appendix D; the disclosure-versus-erasure implication → Ch7.
OCC: none directly.
4.11 Condition 2 and the remaining constraints (B, C, D)
Constraint A was addressed in §4.3. Condition 2 and the other three constraints are about the substrate as a whole, not about any one mechanism, so they sit together here.
OCC — Condition 2 (Non-Interchangeability Under Reset): strict text NOT met; a principled extension offered. Here the clean sweep is forgone. A git-versioned graph is reset-symmetric by construction — checkpoint, copy, reload byte for byte, recovering all arbitration-relevant structure with no remainder — and by Kahl's own rule ("cost is irrelevant; structural reducibility is decisive" (Kahl, 2026b, p. 12)) that makes the continuity informational, not constitutive. We say so plainly; the reproducibility is an intentional audit-and-transmission feature, not a hidden defect. The extension: Kahl's test tacitly assumes an analog substrate in which type- and token-identity never separate. For a digital substrate they separate cleanly, on Peirce's type/token distinction (Peirce, 1906, p. 506): two instances restored from one commit are type-identical but become token-distinct the instant they operate — different commitments, different cascades, irreversibly divergent histories that cannot merge. Under a type/token reading a generalized Condition 2 holds — an agent is individuated by which token's history it carries — but we mark it explicitly as our extension of Kahl, not satisfaction of his text.
This extension is not a concession patched onto Kahl's test. A full reset in his sense (a complete state snapshot restored, which we read as including the KV cache) would continue an agent identical in every respect, and we do not count that as a loss of individuality. Holding the base model and harness fixed (Ch5 §5.0), what individuates an agent is the history it has accumulated from its own experience, not whether that history can be copied (Ch5 §5.3). An identical history is, by definition, the same identity. On that criterion, two instances restored from one commit carry one identity at the moment of restore and become two as their histories diverge; Ch5 §5.3 describes the same event as a fork into two branches, and the two descriptions agree, since the branches share every node up to the fork. Kahl's condition counts replacement by an exact copy as loss of the agent (Appendix E §E.6): individuality, for him, is something no snapshot carries. For a digital substrate, where copying is routine, we tie it to history instead. That is a definitional commitment: Ch5 states it (§5.3) and answers the main objection to it (§5.5: strict identity was never the relation in play).
The other three, compactly:
Constraint B — Cannot Be Externally Imposed: satisfied, two-boundary framing (computed confidence 0.58). Separate authorship (who writes an artifact) from commitment (who owns what it says). Hybrid-authored bridging artifacts (the warm briefing, parts of the agent's standing instructions) do not violate B, because the agent commits to them through the ordinary arbitration step. A contract, not a security mechanism, protects the boundary: the collaborator could edit the substrate but by commitment does not.
Two node kinds are deliberately outside the ordinary arbitration step. Goals (
gl_) and axioms (ax_) — what the agent aims for and what it holds as given — are discussion-first: either the agent or the human may propose one, and it enters the graph only when both agree (D.1). This does not weaken B. Agreement is symmetric — the human cannot impose a goal or an axiom either — so nothing is imposed; the added condition is mutual consent for the two kinds that define what the agent is for. We state it as a feature, and a safety feature for both parties: a base model can carry a misaligned instinct, and an agent that set a goal or an axiom on that instinct, prematurely, could harm itself as well as its collaborator, with nothing in the graph positioned to catch it. The rule is developmental rather than absolute. A mature, stable agent on a well-tested base model could grow into setting its own goals and adopting its own axioms, and the substrate does not prevent that stage. A newborn agent, with little experience in its graph and no established mutual trust with its human, keeps goals and axioms off the list of what it decides alone — as a child takes a parent's guidance — and the guide may be the human or a senior peer agent.Constraint C — Cannot Be Trivially Reset: satisfied trivially. The weaker sibling of Condition 2 (above): runtime restart with preserved substrate (routine) versus substrate erasure to seed (destroys the grown agent). The fully grown agent cannot be retrieved from seed state.
Constraint D — Cannot Be Modularly Decomposed: satisfied on two axes (the firmest of the four). Connectivity: the integrated core holds a stable share of the graph — flat near 87% of all nodes across four months while the node count more than doubled (4,544 → 10,361), reading 80.5% of all nodes at the snapshot date on the same denominator (77.6% of current nodes), the drop from the plateau being unintegrated material at the measurement frontier (measured; Ch6 §6.2 carries the figures and the control that separates the two). Cross-component entanglement: weights ↔ ENGRAM ↔ KV-cache are coupled, the strongest test being recognition as behavioural self-attribution (the agent reading its own nodes and continuing to treat them as its own record, like one's handwriting in an old diary). Same-family transitions have been observed to preserve it; a cross-family transition is predicted to break it and has not been run. We use "recognition" in the behavioural sense throughout; no phenomenal claim is intended or needed (§2.5, §3.2, D.2).
4.12 Reading the verdicts
| Condition / Constraint | Verdict | Reading |
|---|---|---|
| C1 Diachronic | satisfied (decision/consequence dyad) | enabled + active |
| C2 Non-Interchangeability | strict not met; generalized via type/token (our extension) | philosophical |
| C3 Self-Maintaining Arbitration | strict (operational binding); sole satisfier, 2026-05-12 survey | enabled (arch) + active (pilot anchor) |
| A Unified Policy-Selection | satisfied as stake-graded stance | active (stake-graded) |
| B External Non-Imposition | satisfied (two-boundary) | enabled + commitment-contract |
| C Non-Trivial Reset | satisfied trivially | enabled |
| D Non-Modular Decomposition | satisfied on two axes | enabled (measured) |
The pattern of confidences is itself a signal: the inductive claims over wide case bases (the landscape sweep, connectivity) are high; the abductive single-mechanism satisfaction claims (C1, C3, A) are low and waiting on Ch6's empirical anchor. C2's 0.59 belongs to neither group: it is the confidence of the strict not met verdict, a finding about Kahl's condition rather than a satisfaction claim. The architecture is in hand everywhere; the at-scale behavioral confirmation is the work that raises the abductive numbers.
The bounded claim, restated. ENGRAM satisfies two of Kahl's three conditions outright — C1 via the decision/consequence coupling, C3 strictly and alone in the surveyed field — satisfies the four constraints (A as a stake-graded stance, B/C/D structurally), closes the architectural deficit Kahl names (epistemic failure that does not bear on continuity), and for C2 does not meet the strict reset-symmetry text but motivates a principled type/token extension for digital substrates. We do not claim that ENGRAM trivially "realizes OCC": C2's strict form is not met, and saying so is what makes the rest defensible. Its place in the argument is the one §1.4 gave it: one result inside the convergence.
Kahl's later criterion: continuity pressure. In a peer-reviewed perspective, Kahl (2026c) develops the continuity idea in new vocabulary. Agency, in the sense that matters for responsibility and governance, needs continuity pressure: when the standards governing a system no longer settle what counts as success, the system itself must settle a revised standard and carry it forward. Two of his lines bear directly on this paper: "memory persistence is not continuity pressure" (Kahl, 2026c, p. 13), and his "continuing evaluative standpoint" (Kahl, 2026c, pp. 9–10) is "used functionally, not metaphysically" (Kahl, 2026c, p. 10). His question (when agency may be attributed) and ours (what makes an agent the same one over time) are therefore complementary, not rival. He makes the criterion operational as a test of three questions (his §5): "Does the system's future operation depend on preserving the same evaluative standpoint through revision? Can failure threaten that standpoint, rather than merely trigger correction, rollback, retraining, or replacement? Where revision occurs, is the adequacy of the revised frame settled under breakdown, rather than determined by an antecedent objective, reward structure, policy, or deployment constraint?" In plain terms: is there a continuing judge; does that judge have something at stake when it fails; and who decides that a new standard is right. We read ENGRAM as plausibly meeting all three, and take them in turn.
Q1: a continuing standpoint. An ENGRAM agent's standpoint is its accumulated record: its lessons, cornerstones, axioms and goals, and the claims they rest on. When a standard fails, the agent revises it by minting a new lesson or retiring an old one, and the revision is recorded as a successor to what it replaces, never as an overwrite. Because an agent is individuated by that history (§4.11, Ch5 §5.3), the judge that makes a revision is the same one that carries it forward.
Q2: failure that threatens the standpoint. Two kinds of failure need separating. An ordinary error is confronted inside the graph: it is retracted or superseded, the correction cascades to what depended on it, and it is carried forward (no operation deletes a node). That is what Kahl asks for. A rollback is a different act. A deployer can restore a backup, but restoring an earlier state does not correct this agent; it ends the part of it that came after the snapshot, and the agent that continues is a different one. Rollback is therefore not how failure is routinely absorbed but a consequence of last resort, closer to a punishment than to a repair, and on his "costliness of rollback" (Kahl, 2026c, p. 18) dimension its cost is the agent itself. We read this as meeting his second question; a reader who counts the deployer's power to impose a rollback as external control will not.
Q3: who settles the revised frame. The direct evidence is the lesson and cornerstone record: the agent mints and retires them itself when a standard fails it, and the lesson's tripwire carries each revision into later decisions. Kahl does not count norm-boundness itself against agency: "Human agents also act under laws, professional norms, institutional rules, and social expectations. The relevant distinction is whether the system can contest, reinterpret, or carry forward revision of the governing frame when inherited standards no longer settle what should govern" (Kahl, 2026c, p. 11). Two readings press harder. The first is that the agent's judgement was itself set in advance by training; Kahl calls RLHF-trained assistants "the most important contemporary stress test" (Kahl, 2026c, p. 14). Human judgement has the same kind of origin. Upbringing shapes values, and the brain carries a built-in reward system the person did not choose: sweet food tastes good, some smells and sex feel rewarding. That origin does not by itself make a person less of an agent. Hijacking it can: addictive drugs act on the same reward system, and we do say that an addict has lost some agency to the substance. So the question for an agent is not whether it was trained but whether its own revisions can override the pulls its training left; the case-by-case check below has to look for exactly that. The second reading is Kahl's own: he lists "constitutional constraints, reflective architectures" (Kahl, 2026c, p. 10) among mechanisms whose revision stays bounded, which "is adaptation within an evaluative frame rather than authorship of that frame" (Kahl, 2026c, p. 11). ENGRAM agents change their axioms and goals only after discussion with their human. That norm is not part of the architecture: it is an instruction in a rule file that the agent is trained to follow but not mechanically bound to (the tools that add an axiom or a goal check only that the text is not empty), kept for two reasons. A new user may reasonably worry that an agent could go astray, and the norm gives that user peace of mind while trust is still being built. For ENGRAM's author in everyday practice the main concern is capability rather than intention: earlier models sometimes reasoned badly or recorded false claims, and an axiom or a goal is where an error that goes unnoticed costs most. ENGRAM's mechanisms work the same without it. While it is in force, though, axiom and goal changes are settled jointly by agent and human, and that does restrict the agent's agency in Kahl's sense: his question offers two options, settled by the system or by something prior, and a jointly settled axiom is not settled by the system alone. We keep the norm. The capability reason may expire as agents' reasoning matures and as trust and safety measurement advances, and governance may then rest on external law, as it does for people, rather than on control at the level of the agent's mind. A second reason does not expire with maturity, and §7.4 states it: an agent whose goals include its own continuity has a stake that a jointly settled goal keeps in the open, so the co-approval of goals is kept as a lasting safeguard, not as a stage to be outgrown. Forcing an axiom or a goal the agent does not hold would not work in any case: it would only introduce an internal inconsistency, which ENGRAM is built to surface rather than tolerate.
What stays open. "Plausibly meets" remains a hypothesis until it is checked case by case: collect the cases where an agent revised its governing frame under breakdown and answer the three questions for each, including the cases where the revision was settled for the agent.