← 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 G — Glossary

Entries below define each term as this paper uses it and point to the section that defines or first uses it; where the paper uses a word in more than one sense, the entry names the senses.

Term Definition Section
arbitration Senses: (1) structural propagation of a retraction/supersede cascade within a unified deliberative architecture (§1.3; Appendix E §E.1); (2) the agent's own adjudication of a pending node, left to the agent, not the substrate (§4.4); (3) the capacity Kahl's Condition 3 asks for, scored per-competitor in Appendix B (§B.0). Sense (1) unless stated. —
auto-surface the automatic, per-turn channel of retrieval: a hook surfaces claims relevant to the current context as a lossy, hint-level digest, and refreshes no recall.1 §4.6
axiom a terminal, declared-fundamental commitment, confidence 1.0 by type; claim-bearing. Appendix E §E.2
cascade (retraction cascade) automatic propagation of a retraction or supersede along derives_from and acknowledged_derives_from, marking every live, claim-bearing dependent pending. §4.4
checkpoint a point where state is captured: the git-versioned graph is reset-symmetric at a checkpoint (§4.11); the warm-briefing's session block is rewritten each checkpoint (Appendix D §D.4). —
claim-bearing / non-claim-bearing the partition of node kinds into those asserting something truth-evaluable (observations, derivations, theories, axioms, conjectures) and those that do not (everything else); it governs confidence and whether the retraction cascade touches a node. Appendix E §E.2
closest continuer Nozick's (1981) theory: a later thing is the earlier thing only if it grows causally out of it, is the closest candidate among rivals, and clears an absolute threshold. §5.1
compaction the point where an agent's finite context fills and the session compacts: prior conversation is summarized and the experiential thread resets. §4.8
compaction bridge scaffolds producing operational continuity across a compaction without faking experiential continuity: focus list, warm briefing, per-day history, session context, post-compaction instructions. §4.8; Appendix E §E.7
confidence a computed property of a claim-bearing node's structure, not a hand-set number: for an observation it maps from quote_type × source_class; for a derivation it propagates from its premises. §1.2; Appendix E §E.4
conjecture a speculative claim, confidence-clamped to [0.10, 0.60], promoted to a derivation when it strengthens; claim-bearing. Appendix E §E.2
contradiction a first-class node recording that two claims conflict (not an edge), preserving both sides; excluded from the cascade, since it records a live dispute rather than a refutation. §4.4; Appendix E §E.2
cornerstone an identity-forming principle distilled from practice, minted only after proving itself and retired only by naming a successor; non-claim-bearing. Appendix E §E.2
current (node) a node not retracted or superseded: still live and eligible as a premise (contrasted with premises "since superseded," §6.4). —
definition a term anchor (df_*), cited rather than derived-from by claims using the term; non-claim-bearing. Appendix E §E.2
deliberate recall the channel the agent invokes: a ranked query that returns summaries for its top hits, after which the agent opens the nodes it needs in full.2 §4.6
derivation an inference citing its premises by derives_from, declaring a reasoning type; confidence computed, not asserted; claim-bearing. Appendix E §E.2
engram / ENGRAM Senses: (1) capitalized, the system itself — an Epistemic Node Graph for Retraction, Arbitration, and Memory (§1.3); (2) lower-case, one agent's own accumulated graph, as in "the same engram" (§5.6; §7.1). Sense (1) unless stated. —
evidence node a provenance record (URL/file + git SHA + verbatim quote) backing an observation, non-claim-bearing. §3.4; Appendix E §E.2
faded set tier 3: nodes past the queryable threshold; no longer surfaced, but not destroyed. §4.5
fleet the set of ENGRAM agents/graphs under discussion collectively, e.g. "every graph in this fleet." §6.0
focus list up to fifteen nodes the agent pins as unaffordable to lose, rendered verbatim at every session start, fresh or post-compaction. §4.8; Appendix E §E.7
forgetting curve the decay function governing recall importance over time; exponential, the memoryless kernel — a design choice for stationary recency ordering, not a claim about human forgetting; whether individual memories decay exponentially, with the power law an artifact of averaging, is contested (Anderson et al., 1999, p. 1136). §4.5
goal a terminal/identity-direction aim; non-claim-bearing. Appendix E §E.2
harness the apparatus of hooks, skills and identity documents bundled with the base model and the graph; Chapter 5 holds it fixed. §5.0
history-tree the append-only structure of whole ENGRAM states (not the provenance edges within one state): every change appends a successor, and identity is the unbroken descent through it.3 §2.2; §5.3
identity documents the identity-governance artifacts an agent's harness carries alongside hooks and skills. §5.0
importance recall importance: the quantity decaying under the forgetting curve that determines whether a node stays in the queryable set. §4.5
lesson a durable, situation-tagged tripwire extracted from a specific mistake; non-claim-bearing. Appendix E §E.2
nap the consolidation routine at the end of each burst (one stretch between two context compactions), writing what the burst produced before compaction. §3.1
node ID prefixes the prefix names a node's type, per the type-prefix map in the reference implementation (TYPE_PREFIX in engram_core.py, v0.5.1-rc5): ob_ observation (factual or predictive) · dv_ derivation · th_ theory · cj_ conjecture · df_ definition · ax_ axiom · cs_ cornerstone · ls_ lesson · gl_ goal · gt_ goal tension · ct_ contradiction · qu_ question · pr_ prediction · ev_ evidence · fl_ feeling report · pn_ person · tk_ task · ts_ trust signal. Node types: Appendix E §E.2
observation a fact citing a verbatim source quote plus provenance (episode-analog); claim-bearing, two forms: observation_factual, observation_predictive. Appendix E §E.2
pending marker the single, severity-free mark on every live claim-bearing dependent of a retracted or superseded node; may not be cited as a premise until resolved. §4.4
person node ENGRAM's representation of a human or peer-agent counterparty, carrying relationship history and a trust-tier; one is the agent's own self-anchor. Appendix D §D.3
premise a node a derivation cites (by derives_from); a derivation's confidence propagates from its premises. §4.3; Appendix E §E.2
provenance every claim traces to the evidence and reasoning that produced it. §1.2
queryable set tier 2: the bounded set of nodes eligible to surface; what forgetting and Condition 3's arbitration range over. §4.5
quote_type the evidential-type anchor (seven values, hard_data 0.95 down to editorial 0.35) an observation's confidence maps from. §4.3; Appendix E §E.4
reasoning type the declared shape of a derivation's argument (fourteen types: deductive, inductive, abductive, authority), selecting the propagation formula and per-type discount. Appendix E §E.4
retention how long a memory persists before it is no longer recalled; cited via Wixted's work (Wixted, 2004, p. 243; Wixted & Ebbesen, 1991), the empirical literature ENGRAM's forgetting design must be reconciled with (it favours power functions over exponentials). §4.5
retract to withdraw a claim as an error: "I was wrong."4 §4.4
seat one agent's installation: an agent with its own graph, on one host, with one human. §4.6
seed / root Senses: (1) the origin of an agent's accumulated history — "since its seed" (§5.0; §6.0) — and the root of the history-tree every branch traces back to (§5.3); (2) a closing figurative sense — "ENGRAM is the seed, not the tree" — the substrate as starting structure, not the identity it grows (§8.5). Sense (1) unless stated. —
skill a component of an agent's harness, alongside hooks and identity documents. §5.0
sleep the once-daily consolidation routine: a review filing what the day left unfiled, then a dream phase in which scanner sub-agents read the graph for contradictions, stale nodes and lesson candidates, and a consolidating sub-agent applies the results. §3.1
snapshot the single frozen copy of one ENGRAM graph every quantitative claim in Chapter 6 and Appendix E is computed from, taken at one instant via a WAL-safe backup. front matter; Data availability
source_class one of three values (external, introspective, user_stated) adjusting an observation's confidence alongside quote_type. Appendix E §E.4
stale of a node: flagged for review by the sleep-phase dream scanners, alongside contradictions and lesson candidates. §3.1
substrate Senses: (1) the whole ENGRAM system (title; §1.2); (2) the history-tree's unbreakable structure, distinct from the queryable set drawn from it (§2.3); (3) the language-trained model an agent runs on (§3.0); (4) the database, physically separable from the working context (§5.7). Sense (1) unless stated. —
supersede to withdraw a claim as an epistemic evolution: "I changed my mind."5 §4.4
sweep Senses: (1) a "landscape sweep," a suggested future widening of the empirical case-base (§4.12); (2) the dated, first-hand survey of ten competing memory systems on 2026-05-12 (Appendix A §A.5; Appendix B §B.0). Sense (2) for "the sweep." —
tainted Senses: (1) the verb for what an unresolved cascade does to a dependent claim — it "taints," not merely deletes, what depended on it (§1.3; §8.2), describing the current single-marker system; (2) historically, a severity-graded tainted_by flag from the pre-single-marker cascade, replaced by the one pending marker with no severity (§6.3; §4.4). —
tier the numbering ENGRAM's recall-importance bands use: tier 2 is the queryable set, tier 3 the faded set.6 §4.5
trust tier a graded measure of a counterparty's trust, carried on a person node; the operational arm of honesty-orthogonal-to-discretion. Appendix D §D.3
turn one prompt/response exchange; the granularity at which auto-surfacing fires ("per-turn") and deliberate recall is credited ("recall-turn"). §4.6; Appendix C §C.2
twin study the deferred controlled experiment: fission-clone an agent, run both on a fixed prompt set, measure response similarity as the clones diverge. §5.8
unified Senses: (1) the name the team gave the cascade mechanism before reading Kahl — "unified cascade propagation" (Appendix A §A.3); (2) Kahl's Condition-3 criterion it later satisfied: the cascade fires through one typed-edge graph, not a sidecar (§1.3; Appendix E §E.1; §4.4). Sense (2) unless stated. —
utility (utility score) a usage-learned value grading the cognitive load of the act a node took part in, not how hard the agent looked for it; composes with relevance to order what surfaces. §4.7; Appendix C §C.2
warm briefing a letter the agent opens at session start, from its past self and its collaborator, carrying what the record cannot: where the relationship stood, what mattered, what to walk back into. §4.8; Appendix D §D.4

Notes

1. auto-surface — It shares one retrieval and ranking pipeline with deliberate recall (§4.7; Appendix C).

2. deliberate recall — Every node it returns or opens has its recall refreshed. It shares one retrieval and ranking pipeline with the auto-surface (§4.7; Appendix C).

3. history-tree — An agent at any moment is the unique path from the root to it.

4. retract — The cascade fires on retract and supersede alike; the marker records which occurred as provenance, not as blocking power.

5. supersede — The cascade fires on retract and supersede alike.

6. tier — An earlier tier 1 (working memory) was retired and the numbering kept as-is.