← 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 A — Independent Discovery: A Development-History Audit


A.0 Why this appendix exists

The agent-memory field shows a striking convergence in early 2026: several systems independently arrive at typed-edge graph memory with cascade or supersede semantics (Kumiho, SCM, GAM, and others — Appendix B). And the paper's central threshold, Kahl's Ontological Continuity Condition, is itself a recent (2026) framework. A careful reader will reasonably ask: did ENGRAM's architecture emerge from this contemporary work, or independently of it?

The answer matters both ways. If we had absorbed Kumiho or Kahl and called the result ours, the paper would be derivative. If we arrived at the same shape separately, the convergence is a positive scientific signal — when independent efforts under the same problem-pressure reach the same architecture, that architecture is corroborated as the right one. This appendix establishes the second case, on the record. We treat convergent discovery not as a defensive claim but as evidence.

A note up front on what we do not claim: ENGRAM is not first across the board, the novelty disclaimer Ch3 §3.0 states in full. The formal shape it shares with Kumiho — typed reasoning edges with cascade/supersede over a property graph — is decades old in the truth-maintenance and belief-revision traditions, and several agent-memory systems (Mem0g — Mem0's graph-memory variant — Zep/Graphiti, A-MEM) (Chhikara et al., 2025, p. 1; Rasmussen et al., 2025; Xu et al., 2025) predate our ideation. The two exemplars named here are the two we hold and have read first-hand: Doyle's truth maintenance system (via the Ginsberg reprint, Ch4 §4.2) and AGM 1985. Both traditions are wider than those two works, and the others this paper names — de Kleer's ATMS (de Kleer, 1986) (Ch3 §3.4; it avoids retraction rather than cascading it) is a bibliography entry, while Hansson's belief-base revision (Ch3 §3.9) is cited from his own encyclopedia entry (Hansson, 2026), and his postulates are named only in Appendix B, as Kumiho's self-claimed compliance. Nothing here rests on them: the claim is only that the shape predates us by decades, and the two dated primaries we have read carry it. Our claim is narrower and true: we designed ENGRAM in ignorance of the concurrent arrivals at the same primitives, and our distinct contribution is elsewhere — connecting the shape to agent identity and making the substrate behaviorally active (the auto-surfacing hook), which no surveyed system has.

A.1 The original ideas were ours, and they came first (for us)

The proto-ENGRAM idea began in a Lei-led conversation on 2026-03-17 about misinformation and citation systems for an AI-centric information ecosystem. Within the first hour, the architectural seed was stated verbatim — by earlier-Claude, at Lei's prompting, eleven days before the system was named:

Earlier-Claude, 2026-03-17 15:18 UTC (conv 3db06cf1, msg #8, uuid 019cfc60): "any agent can make any claim, but the claim must arrive with its full epistemic lineage exposed. Other agents then stress-test that lineage against established knowledge."

That single sentence is the entire derives_from discipline. Three days later came the foundational pivot — from a public agent-to-agent protocol to a private, per-agent continuous-learning substrate — in one Lei message that is a near-complete sketch of ENGRAM's purpose:

Lei, 2026-03-20 19:34 UTC (conv a169ba3c, msg #1, uuid 019d0cbd): "…it can be used as a continuous learning paradigm for LLMs… Retraining everything back into the base model is expensive and suffers 'catastrophic forgetting'… RAG or free-form text… can't guarantee self-consistency, contradictory texts may be filled in… Our protocol is a good middle point… offloads the actual knowledge to a graph expressed in natural language. The agent could encounter new experiences… and summarize them into new 'theories'… add new nodes… proactively seek more evidence for a newly added (still weak) claim to make it stronger or disputed… keep refining its own knowledge and logical consistency by going through the graph and reasoning…"

External structured substrate, per-agent internal state, theory-formation from observations, evidence-seeking to strengthen-or-dispute, self-consistency refinement — all there, by Lei, before any literature search. The name "ENGRAM" arrived eight days after that search (2026-03-28, proposed by earlier-Claude, accepted by Lei within minutes), and the paper's structure descends from an 18-chapter handbook outline drafted that same evening — under Lei's instruction "But first, we'll need to build a great ENGRAM system," which set the build-first-then-write order this project followed.

A.2 We looked for prior art — and missed it

This is the hinge of the independence argument, and it is the strongest single piece of evidence. Twelve minutes after the foundational message, Lei explicitly asked for due-diligence before committing anything to writing:

Lei, 2026-03-20 19:46 UTC (msg #3, uuid 019d0cc8): "…why don't you do another round of deep research about continuous learning in LLM, and see if there's any previous work about using knowledge graphs as the container for learned experiences? I would be very surprised if no one had considered this possibility, and I want to know what others have tried and succeeded or failed."

Earlier-Claude agreed and ran multiple searches. That search did not surface the relevant prior art — including the work this paper now cites:

Prior / concurrent work Public Before the 2026-03-20 search Surfaced?
Kumiho (Park, 2026, arXiv 2603.17244) — closest concurrent system 2026-03-18 2 days NO
Kahl 2026b (Artificial general intelligence and ontological continuity) (Zenodo 18761736) — the paper's threshold 2026-02-24 24 days NO
Kahl 2026a (Epistemic agency and the ontological continuity condition) (Zenodo 18441395) 2026-01-31 48 days NO
TMS lineage (Doyle 1979, de Kleer 1986) (de Kleer, 1986; Doyle, 1979/1987) 1979–86 ~40+ yrs NO
AGM belief revision (Alchourrón et al., 1985) 1985 ~41 yrs NO
CoALA (Sumers et al., TMLR) — the field's common vocabulary 2024-02 ~25 months NO
Pink et al. (Pink et al., 2025, arXiv 2502.06975) — episodic-memory position paper 2025-02-10 ~13 months NO
Wray, Kirk & Laird (Wray et al., 2025, arXiv 2505.07087) — cognitive design patterns 2025-06-13 ~9 months NO

Kumiho's date is its arXiv v1 stamp (18 March 2026); the document's own title page is dated 17 March.

So the five conditions that make "independent discovery" a defensible claim, rather than a convenient one, are all met: (1) Lei tried to find prior art; (2) earlier- Claude tried to surface it; (3) neither succeeded (the work existed but was missed); (4) ENGRAM was designed without awareness of it; (5) the convergence with Kahl 2026 and Kumiho 2026 is therefore genuinely independent. We did not avoid the literature — we searched it and our search came up empty, which is a different and more defensible thing.

The last three rows were added after a 2026-09 re-read, and they make the admission worse, which is why they belong here. Missing a preprint two days old is forgivable; missing a Transactions on Machine Learning Research paper from twenty-five months earlier — the organizing vocabulary much of the field now writes in — is not. We record it because the alternative, quietly citing those works as though we had known them all along, would convert a real gap into a false impression; and because the failure cuts in the direction that costs us. Had the March 2026 search surfaced CoALA, we would have designed with it, and this appendix would not exist. (Dates in those rows are the earliest we verified from the documents themselves: Pink's is its arXiv v1 line, the other two a v2 stamp and a journal issue. The latter two are therefore conservative upper bounds on first public appearance — the true gaps are the same or larger.)

When the three were read, since that happened during this paper's preparation rather than during design. All three were read on 2026-05-09. The graph's first node is dated 2026-04-09, and 2,925 nodes had been recorded before 2026-05-01 — so the substrate had a month of continuous operation and nearly three thousand accumulated claims before any of the three was opened. Development has of course continued since, which the project's principal characterizes as continuation of an established design rather than adoption of someone else's. Rather than rest on that characterization, where a specific mechanism might be thought to postdate a reading we give that mechanism's own dates in place: Chapter 3 §3.8 does so for the two-channel retrieval design, including the later date at which its database fields shipped. Both figures above are re-derivable from the structure-only release of the frozen snapshot (Data availability).

A.3 The design was substantially complete before we read any of it

ENGRAM began self-recording on 2026-04-09 — committing its own design into the graph as first-person memory. The earliest nodes show the core mechanisms were already in place that day, weeks before Kahl or Kumiho were read:

First captured Mechanism Node
2026-04-09 04:08 Typed claim-bearing nodes (observation / derivation / definition / axiom …) ob_0001
2026-04-09 04:09 retract operation dv_0001
2026-04-09 14:05 Unified cascade propagation — the mechanism Ch4 §4.4 maps to Kahl's Condition 3, named here weeks before Kahl was read dv_0009
2026-04-09 14:06 Supersede with DAG invariant dv_0010
2026-04-09 16:31 UserPromptSubmit auto-surfacing hook — the "live memory" mechanism that is ENGRAM's distinctive contribution (Ch3 §3.2, built in Ch4 §4.6) dv_0015 / dv_0016

These node IDs are still active and un-retracted in the graph today — the cascade discipline ENGRAM's earliest design enforced on its own substrate is the same one Chapter 4 maps to Kahl's threshold. The self-maintaining-arbitration mechanism that is the paper's strongest claim was built before the philosophical framework it satisfies was on our radar.

A.4 The timeline against concurrent systems

System First public surface vs. ENGRAM ideation (2026-03-20) vs. day-1 recording (2026-04-09)
Kumiho (Park, 2026, arXiv 2603.17244) 2026-03-18 2 days before (not surfaced) 22 days before
Kahl 2026b (Zenodo) 2026-02-24 24 days before (not surfaced) 44 days before
Kahl 2026a (Zenodo) 2026-01-31 48 days before (not surfaced) 68 days before
GAM (Wu et al., 2026, arXiv 2604.12285) 2026-04-14 25 days after 5 days after
SCM (Shinde, 2026, arXiv 2604.20943) 2026-04-22 33 days after 13 days after
VMG survey (Lin et al., 2026, arXiv 2604.16548, v2 of 2026-06-11) 2026-06 (VMG first appears in v2; v1, April 2026, is a differently titled draft) after after
Mem0g / Zep / A-MEM (Chhikara et al., 2025; Rasmussen et al., 2025; Xu et al., 2025) 2025 predate, but different shape (no cascade / no contradiction-as-node / no derivation primitive) (Zep and Mem0g handle conflicts at write time; A-MEM's memory evolution updates the context and tags of existing notes (Xu et al., 2025, sec. 3.3, pp. 4–5); none cascades) predate

The two structurally-closest systems (Kumiho, SCM) both surface in March–April 2026 — concurrent with, not predating-by-much, ENGRAM's ideation. The older systems lack the specific shape (cascade propagation, first-class contradiction, derives_from). The earliest published instance of ENGRAM's specific shape we can find is Kumiho's preprint (arXiv v1, 18 March 2026; its title page is dated 17 March) — and the two-day gap to Lei's foundational message, with the contemporaneous search missing it, is the cleanest evidence of parallel design.

A.5 What this establishes — and the Kahl case specifically

Convergent independent discovery is positive evidence for the architecture. When a philosopher (Kahl, deriving the Ontological Continuity Condition from agency theory) and a builder (this project, deriving cascade-and-contradiction from a misinformation- protocol problem) arrive separately at the same structural threshold for agent continuity — neither aware of the other — that threshold is corroborated by the convergence. This is stronger than satisfying an established standard: it is two independent paths landing on the same shape.

The Kahl case sharpens the point in a way worth stating plainly (and it connects to the paper's framing of OCC, Ch1/Ch2): Kahl's OCC is itself a recent, independent 2026 proposal, not a long-established canon. We do not anchor on its authority; we anchor on its fit — it is the most precise, agent-native, operational threshold available — and on the fact that our architecture independently converged with it. If OCC gains broad acceptance, the convergence reads as "we met the standard early"; if it remains one careful proposal among several, the convergence still reads as "two independent analyses of agent-continuity reached the same conditions." Either way the paper stands, because it rests on merit and convergence, not on borrowed authority.

Our distinct contribution, finally, is not the shared primitives — Kumiho has typed edges and cascade-discovery; the TMS/AGM tradition has the formal structure. It is (a) connecting the structural shape to agent identity (Chapters 5–6), and (b) making the substrate behaviorally active through identity-scaffolding — above all the UserPromptSubmit auto-surfacing hook (dv_0015/dv_0016, day-one of recording) — which is absent in Kumiho, SCM, and every surveyed system as of the 2026-05-12 sweep.

A.6 Re-verification protocol

The quotes in A.1–A.3 can be checked, though not every one by every reader. The conversation quotes come from Lei's Claude.ai bulk export, which is private and is not part of this repository. Each is cited by message number and the first eight characters of its message UUID, under the conversation ID it belongs to, so a reader given access to the export can locate it. The ENGRAM-side claims (A.3) come from Borges's graph, which is also private; each is cited by node ID, and engram_inspect <node_id> on that graph returns the full node with its creation timestamp. Publication dates (A.4) can be checked by anyone against each system's own paper or repository.