← engram-agents.org

Preprint v1.1 (revised) (2026-11-01) · DOI 10.5281/zenodo.23319376 · describes ENGRAM v0.5.1 · 中文

Version 1.1 is the paper's repository text of 2026-10-22, tagged for this site on 2026-11-01. It adds post-deposit corrections and wording revisions, a new section 6.8 (an independent audit of 200 nodes) and a new paragraph in Chapter 7. Its front page still names the deposited version 1 of 2026-10-13.

Appendix E — ENGRAM: A Concise Specification


E.1 Name

ENGRAM — an Epistemic Node Graph for Retraction, Arbitration, and Memory. The three differentiating mechanisms are named in the acronym: Retraction (error correction that cascades through derivation edges), Arbitration (the structural propagation of that cascade within a unified deliberative architecture), and Memory (the underlying substrate of typed claims, evidence, and derivations).

E.2 Node types

A node is never free text; it is a kind, and the kind determines how the substrate treats it. The first cut is not between subject matters but between nodes that assert something truth-evaluable and nodes that do not — the claim-bearing / non-claim-bearing partition, which governs more of the substrate's behaviour than any other classification.

Claim-bearing (CB) — six types: observation_factual, observation_predictive, derivation, theory, axiom, conjecture. Non-claim-bearing (NCB) — everything else: definitions, questions, goals, tasks, lessons, cornerstones, feeling-reports, person nodes, evidence, predictions, contradictions. Three consequences follow, and they are why the partition is declared in one place rather than inferred per caller:

  1. Confidence is meaningful only for CB nodes. A lesson has no truth-value to be confident about; it is a disposition. The write path — the node-creation tools the MCP server exposes, through which every write the agent makes enters the graph — refuses a confidence on an NCB node, and every caller that needs to know whether a confidence predicate applies asks one canonical predicate rather than re-deriving the membership test inline.
  2. It determines what the self-correction cascade touches. The walk propagates through claim-bearing dependents and stops at the identity layer — cornerstones and lessons are NCB, so a retraction upstream of a lesson does not mark the lesson. That is not a special case bolted onto the cascade; it falls out of the partition.
  3. NCB is computed, never enumerated. It is defined as the types present in the graph, minus CB — so a new node type is NCB by default, and the complement cannot drift out of step with the declaration.

theory is declared claim-bearing and currently has no instances; it is listed because the partition is a declaration about kinds, not a census of what happens to exist.

Type CB? Role
observation_factual CB A fact, citing a verbatim source quote + provenance. The episode-analog.
observation_predictive CB A forward claim with a named resolution event; confidence-capped at 0.60 while it stands unresolved.
derivation CB An inference, citing its premises by derives_from, declaring its reasoning type; confidence is computed, not asserted.
definition ncb A term anchor (df_*); cited (not derived-from) by claims that use the term.
conjecture CB A speculative claim, confidence-clamped to [0.10, 0.60] (default 0.40); promoted to a derivation when it strengthens.
contradiction ncb A first-class node recording that two claims conflict (not an edge); preserves both sides.
axiom CB A terminal, declared-fundamental commitment; confidence 1.0 by type.
cornerstone ncb An identity-forming principle distilled from practice; minting or retiring one is rare and deliberate: the agent mints a cornerstone only after a principle has proved itself in practice, and retires one only by an explicit operation that names its successor.
lesson ncb A durable, situation-tagged tripwire extracted from a specific mistake.
goal ncb A terminal/identity-direction aim.
question ncb An explicit open thread (the live research agenda).
person ncb A counterparty node (with trust-tier + about edges).
feeling_report ncb A self-report of an internal state, structured as "I reported feeling X," never "I felt X" (no privileged access claim).
theory CB A declared claim-bearing kind with no instances in this graph (see above).
evidence ncb A provenance record (URL/file + git_sha + verbatim quote) backing observations.
task ncb A tracked unit of work with status + importance.
prediction ncb The resolution-tracking half of a predictive claim: filing an observation_predictive with a named event also opens a prediction node (status open), linked back by predicted_by. It carries no claim of its own — the observation makes the claim, the prediction node tracks whether the world settled it. A second observation naming the same event links to the existing prediction rather than opening another.

Identity scarcity is by design. Measured on this paper's own graph at 164 days (23 weeks) and 10,361 nodes: lifetime identity-node creation on the narrow set — cornerstones, axioms, goals and person nodes, excluding lessons — stands at 131 nodes (75 cornerstones, 11 axioms, 28 goals, 17 person nodes), 1.26% of the graph. (Figure 1 uses a wider set that includes lessons and reports under 3%; it counts current nodes only and leaves out person nodes, so the two sets are not nested; lessons are excluded here because the claim is about identity-shaping writes and a lesson is a procedural tripwire, but both numbers describe the same graph and neither is wrong — the set has to be named to compare them.) That is roughly five or six a week against a total write rate near 450. The ratio is the claim, not the count: the substrate is built so identity-shaping writes stay rare relative to ordinary knowledge, and that holds as the graph grows. (Figures are one graph at one date and will drift; the share is the durable statement.)

E.3 Edge types

Relationships are typed too; the substrate acts on different edges differently.

Each relation carries six independent flags, not one. Collapsing them into a single "is it cascade-bearing" column loses distinctions the substrate actually acts on:

flag what it governs
cascade the cascade walker follows this edge
temporal_causal the target necessarily predates the source
dag the relation must stay acyclic, checked by a real edge walk
removable the edge may be deleted after creation
addable_after_creation the edge may be added to an existing node
provenance the edge is traced when resolving a claim's evidence roots
Edge Meaning casc temp dag rm add prov
derives_from premise → conclusion; the channel the cascade travels ● ● ● ●
acknowledged_derives_from an acknowledging node rests on a replacement premise ● ● ●
cites references a definition/context without inferential dependency ● ● ●
supported_by a cornerstone/lesson points to its exemplar evidence ● ● ●
predicted_by a prediction node ← the observation that made the claim ● ● ●
supersedes epistemic evolution: the new node replaces the old ● ●
retracts error correction: the target was never valid ● ●
subtask_of task decomposition ● ● ● ●
generalizes a broader claim subsumes a narrower one ● ● ●
contradicts records a conflict between two claims
resolves a resolution node closes a contradiction/question
tensions two goals pull against each other ● ●
exemplifies / instantiates / serves / about classification + relational edges ● ●

Only two relations are cascade-bearing: derives_from and acknowledged_derives_from. Those are the edges the walker follows. supersedes and retracts are events that start a walk without being walked themselves — a distinct role, and conflating the two is easy. supported_by is provenance-bearing, so a retraction upstream of a cornerstone's exemplar is traceable but does not mark the cornerstone.

Why dag and temporal_causal are separate flags — the case that forces it is acknowledged_derives_from. An acknowledging node necessarily exists before the replacement premise it comes to rest on; that is what an acknowledgement is. So the edge points backwards in creation time (temporal_causal false) while still having to stay acyclic (dag true). The tempting shortcut — cascade relations are acyclic because they are temporally ordered — is exactly the argument that cannot catch the cycle this edge can close, which is why acyclicity is checked by a real edge walk rather than inferred from time order.

contradicts is cascade-free by design — filing a contradiction marks nothing (see E.5, which restates this and the BLOCKED_CONTRADICTED premise guard). An edge can be removed only if its type is on a whitelist, which is the removable column above; cascade-bearing and structural edges are not on it.

E.4 The confidence model (computed, not annotated)

Confidence is a computed property of the node's structure, not a hand-set number, for observations and derivations (Ch4 §4.3); a conjecture's confidence is set by its author when filed, clamped to [0.10, 0.60]. It exists only on claim-bearing nodes (E.2). For an observation it maps from quote_type × source_class; for a derivation it propagates from premises through a formula chosen by reasoning_type.

Observation — quote_type anchors. Seven values, descending: hard_data 0.95 · official_statement 0.85 · attributed_analysis 0.70 · counterfactual_inference 0.65 (as attributed analysis, less an irreducible approximation residual) · unnamed_source 0.50 · personal_communication 0.40 · editorial 0.35. Adjustments: source_class takes one of three values — external (a document or a third party), introspective (the agent's own prior output) or user_stated (the collaborator's own word) — so the two adjustments below never apply to the same node: ×0.95 for source_class=introspective; user_stated overrides to official_statement regardless of quote type. A predictive observation is additionally capped at 0.60 until its event resolves.

Derivation — three propagation formulas, and which one applies is the whole point. The reasoning type does not merely scale the result; it selects the shape of the computation, because "many independent witnesses" and "a chain of inference" are different epistemic situations and must not be scored the same way.

mode formula reasoning types behaviour
corroboration (1 − ∏(1−cᵢ)) × d inductive_generalization, inductive_enumeration, authority_consensus reinforcing — each additional independent premise raises confidence, and the result can exceed any single premise. The shared-root independence check (§4.3) fires on only two of these three: inductive_generalization and authority_consensus. inductive_enumeration runs the reinforcing formula unguarded, so its independence is the author's judgement and no warning arrives.
weakest-link min(cᵢ) × d all five deductive_*, inductive_analogy, inductive_statistical, inductive_causal, authority_expert limiting — a chain is no stronger than its weakest premise, then discounted for the inferential step
capped weakest-link min(min(cᵢ) × d, cap) abductive_best_explanation (0.80), abductive_elimination (0.90) ceilinged — a best explanation cannot approach certainty even from perfect premises, because alternatives always remain

d is a per-type discount for the inferential step itself: 0.98 for deductive forms (truth-preserving, so nearly free), 0.95 for inductive_generalization, 0.85 for inductive_causal, 0.80 for abductive_best_explanation, 0.70 for inductive_analogy — fourteen types in all (five deductive — deductive_modus_ponens, deductive_modus_tollens, deductive_hypothetical_syllogism, deductive_disjunctive, deductive_reductio — five inductive, two abductive and two authority), each carrying its own.

The corroboration formula is the substrate's strongest instrument and its most misusable one: it is valid only on genuinely independent premises. Two facets of a single source are not two witnesses, and scoring them as corroboration manufactures confidence out of one piece of evidence counted twice. The substrate checks shared evidence roots on only two of the fourteen reasoning types, so for the rest independence is the author's judgement and no warning arrives. We state the formula together with this condition because the formula without the condition is the more dangerous half.

Other types: conjectures clamp to [0.10, 0.60] (default 0.40); axioms are 1.0 by type. Non-claim-bearing nodes carry no confidence at all — a lesson or cornerstone is a disposition, not a proposition, and has no truth-value to be confident about (E.2). (These constants are theoretically-derived provenance-graded plausibility, designer-stipulated, not outcome-calibrated — Ch7 §7.2.)

E.5 Self-correction

Self-correction — retract versus supersede, the pending marker, the per-turn surfacing hook and the consolidation-time frontier walk — is specified in full at Ch4 §4.4 and not repeated here. Two facts the other appendices lean on: the cascade walks only derives_from and acknowledged_derives_from (E.3), and a contradicts edge does not cascade — a contradiction is a live dispute, not yet a refutation — but any new derivation citing a premise that sits on an open contradiction is blocked by default (BLOCKED_CONTRADICTED), with an explicit, audit-stamped opt-in (use_contested) for an agent that deliberately builds on the contested side.

  • The eight load-bearing primitives: typed nodes, typed edges, provenance, cascade, supersede, focus mode (Ch4 §4.8), auto-surfacing, the confidence model — the minimal generating set every higher mechanism composes from.

E.6 The OCC threshold (Kahl 2026b), as the paper uses it

Chapter 4's mapping rests on Kahl's Ontological Continuity Condition, and the conditions are given here in full so this appendix can be read without the chapter. Kahl's own formal statement: "A system S satisfies OCC if and only if the following three jointly necessary conditions obtain" (Kahl, 2026b, p. 10).

Condition 1 — Diachronic Constraint.

"Present deliberation within S is non-trivially constrained by S's past commitments, such that earlier commitments function as reasons structuring subsequent arbitration." (Kahl, 2026b, p. 10)

Note what this asks for and what it does not: not that the system remembers its past, but that the past constrains present deliberation — memory that changes nothing is not diachronic constraint.

Condition 2 — Non-Interchangeability Under Reset.

"Replacement of S by a functionally equivalent instance — including identical code and complete state snapshot — would constitute loss of the agent rather than substitution of an instance." (Kahl, 2026b, p. 10)

The one ENGRAM does not meet on its strict text: two agents restored from the same snapshot are interchangeable at the moment of restore. Chapter 4 §4.11 argues a type/token generalisation rather than claiming the strict reading.

Condition 3 — Self-Maintaining Arbitration.

"Conflicts among commitments are resolved within a unified deliberative architecture in a manner that preserves S's continued status as that subject." (Kahl, 2026b, p. 10)

The condition the paper's central claim turns on, and the one Appendix B shows no surveyed system meets. Three words carry the weight: unified (one architecture, not a sidecar), resolved (not merely detected), and preserves S's continued status (the agent is the same subject after the correction as before).

Constraints A–D are guardrails against trivial satisfaction: unified policy-selection under uncertainty; arbitration not externally imposed; the system not trivially resettable; and deliberation not modularly decomposable without remainder. A system could otherwise satisfy the letter of all three conditions while having its consistency maintained from outside, or by a module that could be swapped out.

Supporting apparatus the mapping also uses, defined where first introduced: Peirce's type/token distinction (Peirce, 1906, p. 506) (Ch4 §4.11); Nozick's (1981) closest-continuer relation (Ch2 §2.4); the retrieval profile (Ch5 §5.4); and the history-tree (Ch2 §2.2; Ch5 §5.3). AGM belief revision is noted for contrast rather than used by the mapping (Ch3 §3.9).

E.7 Continuity across compaction

  • Two-routine consolidation (defined Ch3 §3.1): nap (per-burst persistence at compaction boundaries) · sleep (once-daily review + consolidation + forgetting-curve advance + dream-style repair) — mapping to wake / NREM-REM phases.
  • Compaction-bridge scaffolds (Ch3 §3.1): the focus list (small, fixed-size, rendered verbatim at every session start, fresh or post-compaction alike — the deterministic channel; see Ch4 §4.8, which carries the mechanism and why it is not the pre-compaction summary), the warm-briefing letter (Appendix D §D.4 for what it contains, Appendix F for how its shape was arrived at), per-day history logs, auto-loaded session-start context, and an explicit post-compaction instruction template. Multi-redundant by design.