← 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 B — The Condition-3 Survey: Why No Surveyed System Self-Maintains Arbitration

B.0 Scope, date, and method

The survey. Eleven sources — ten systems and one survey of the field — read first-hand on 2026-05-12: Pinecone Nexus (Ashutosh & Liberty, 2026), Anthropic's memory tool (Memory Tool, n.d.), Letta (MemGPT) (letta-ai, n.d.; Packer et al., 2023), Mem0 v3 (Mem0, n.d.-a, n.d.-b), Zep/Graphiti (Rasmussen et al., 2025), TierMem (Zhu et al., 2026), MemMachine (Wang et al., 2026a), SCM, GAM, Kumiho, and the 2026 Memory for Autonomous LLM Agents survey (Du, 2026). This is a bounded claim about that surveyed set as of that date. Agent memory is a fast-moving field; we make no claim about systems outside the survey or released after it. Newer or missed systems are exactly why the claim is scoped, dated, and listed rather than universal.

The bar. Kahl's Condition 3 — Self-Maintaining Arbitration, verbatim: "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). In operational terms a system meets C3 only if it automatically resolves contradictions among its own stored claims as an internal state-change (the 'One precision' note below says what 'resolves' does and does not mean) — the cascade propagates by construction and the affected claims surface to the agent without an intentional reach-out. Four ways to fall short:

  • (N) No resolution — contradictions are absent as a construct, or detected but never resolved (the edges/flags persist).
  • (A) Advisory / opt-in — a resolution mechanism exists but must be explicitly invoked, and it recommends rather than retracts.
  • (X) External delegation — resolution is handed to retrieval ranking, recency rules, or application code, outside the deliberative architecture.
  • (S) No claim structure — the system operates on text blocks with nothing to cascade through.

Scoping note. Where a system's C3 failure is shown by a positive design choice (delegate to retrieval, overwrite by recency, advise-don't-retract), we quote it verbatim. Where the source documents detection but is silent on resolution (GAM), there is no sentence to quote — we say so and rest the verdict on the documented absence, not an invented quote.

Synthesis. Across the eleven sources, none satisfies Condition 3 strictly; one of them, TierMem, has no contradiction construct to arbitrate (§B.5). The conclusion is the conjunction of the verdicts above, so it holds exactly as far as each of them does. The field has converged on the structural primitives C3 needs (typed edges, provenance, cascade-traversable graphs, supersede semantics — Kumiho most fully); what is missing everywhere is the operational binding that makes arbitration automatic state-change. ENGRAM's contribution is that binding (Ch4 §4.4), not the primitives.

Section numbers inside paper-verbatim and paper-claimed tags refer to the quoted paper's own sections, not to this paper's.


B.1 The near-miss: a full revision engine, but advisory

Kumiho (arXiv 2603.17244) — (A) advisory / opt-in

The strongest competitor in the survey: typed Depends_On edges, an AnalyzeImpact traversal, and self-claimed AGM K*2–K*6 and Hansson belief-base compliance. The gap is not belief revision — it is that the cascade advises rather than acts.

The compliance claim is theirs, recorded at the fidelity we can vouch for. The AGM and Hansson properties above are Kumiho's own claims about Kumiho, and its paper makes them in two parts — a formal part, "proving satisfaction of the basic AGM postulates (K*2–K*6) and Hansson's belief base postulates (Relevance, Core-Retainment)" (Park, 2026, p. 1), and an implementation part, "Automated AGM compliance testing (49 scenarios, 7 postulates) confirms 100% operational adherence to the claimed formal properties" (Park, 2026, p. 2). That structure is careful rather than confused: the proofs are offered for the semantics, the scenarios for the running system. We can vouch for neither half. We have read their paper; we have not worked through their proofs step by step, and we have no access to their implementation to re-run anything — and 49 passing scenarios would in any case establish that no counterexample was found in 49 cases, not that none exists. So we record the compliance as claimed rather than as established and leave it unadjudicated: nothing in this survey turns on it, because the contrast drawn here is about where a correction binds, not about whose postulates are satisfied. (On why satisfaction is the wrong relation for this paper to assert of itself, see Ch3 §3.9.)

  • paper-verbatim (§6.5): "AnalyzeImpact traverses typed Depends_On edges to identify all downstream conclusions that may need re-evaluation—a capability enabled by the typed edge ontology that flat file systems cannot express."
  • paper-verbatim (edge definition): "Depends_On: Validity dependency. If the target is invalidated, the source may be unreliable." (The modal may be is the advisory tell — not is invalidated.)
  • paper-verbatim (§7.1, consistency enforcement): "deprecated items are excluded from all search and retrieval operations by default" — i.e. consistency is held at the retrieval surface, not by structural arbitration of commitments.
  • first-hand-read finding: the cascade is advisory, not automatic — AnalyzeImpact identifies affected beliefs but does not retract them; a human or agent must review and decide whether to revise. Frame the gap as advisory-vs-automatic, not as "no belief revision" — Kumiho has the revision machinery; it withholds the automatic binding.

B.2 Detect-but-don't-resolve

SCM — Sleep-Consolidated Memory (arXiv 2604.20943) — (N) no resolution

Has a contradicts edge predicate and a sleep cycle, but the resolution is promised, not implemented: the same edges that §3.2 says are "flagged for resolution" are used in §3.5 only as a trigger metric.

  • paper-verbatim (§3.2): "when a user states a preference that contradicts a previously stored preference, the system records a contradicts edge and flags the conflict for resolution during sleep."
  • paper-verbatim (§3.5): "Conflict density is the ratio of contradicts edges to total edges in the long-term memory graph: ρ(G) = |E_contradicts|/|E|."
  • paper-verbatim (§3.5): the sleep cycle is "implementing three distinct offline processes" — NREM consolidation (Hebbian strengthening plus proportional downscaling), REM dreaming (random-walk sequence integration), and value-based forgetting — presented as a state machine (Figure 2) with "the full pseudocode given in Algorithm 1."
  • first-hand-read finding, from the paper alone: none of the three described phases acts on a contradicts edge. The edges appear in the sleep cycle only inside the trigger metric ρ(G), and then persist. The finding is checkable at the level of the whole document: the string resolution occurs twice in the SCM paper — once in §3.2's promise quoted above, and once in future work, listing "contradictory information resolution" among capabilities harder evaluations should test. It occurs nowhere in §3.5, §3.6, or Algorithm 1. Because SCM publishes its sleep cycle's complete state machine and pseudocode, this absence is visible in the paper's own account, and needs no claim about SCM's code. SCM has the surface appearance of structural conflict handling without the substance: no arbitration outcome, no cascade, no downstream taint.

GAM — Hierarchical Graph-based Agentic Memory (arXiv 2604.12285) — (N) detect-only

Assigns a contradict edge at consolidation, but both conflicting nodes persist and nothing arbitrates; conflict is managed by timing, not resolution.

  • paper-verbatim (Appendix E.1, the relation-scoring prompt): "contradict: one memory contradicts the other (asymmetric)" — detection exists as a typed relation.
  • scope: there is no paper sentence stating GAM does not resolve — the C3 failure is shown by the presence of a detect-only edge plus the absence of any resolution prose. First-hand-read finding: "both nodes remain in the [memory]; retrieval ranking implicitly favors higher-confidence information, but no conflict resolution mechanism is documented. There is no arbitration, no automatic retraction, no cascade." We cite this as analysis, not as a quote.

Zep / Graphiti (Rasmussen et al., 2025) — (N) contradiction not persisted, no cascade

Contradiction is an LLM classification made during ingestion that invalidates the directly contradicted edge(s); no contradiction node persists, and entity-edges cannot reference other edges, so there is nothing to cascade through.

  • paper-verbatim (§2.2.3): "The system employs an LLM to compare new edges against semantically related existing edges to identify potential contradictions. When the system identifies temporally overlapping contradictions, it invalidates the affected edges by setting their t_invalid to the t_valid of the invalidating edge."
  • first-hand-read finding: invalidation touches only the directly contradicted edge(s) and is triggered only by new arriving episodes; an agent cannot inspect its own beliefs, find a contradiction, and propagate a retraction. The contradiction is "not persisted as a node, not persisted as an edge type, and not queryable after ingestion."

B.3 Resolve-by-overwrite or by-someone-else (external delegation)

Mem0 v3 (docs.mem0.ai migration guide, as of 2026-05-12) (Mem0, n.d.-a) — (X) external delegation

The clearest external-delegation evidence in the survey. v3 is add-only; contradictions coexist and are sorted out at read time.

  • paper-verbatim: "Retrieval handles ranking — the most relevant, current information surfaces first." (Mem0, n.d.-a, step 4 Tip, Internet Archive snapshot of 2026-05-14)
  • first-hand-read finding: conflict resolution is delegated to retrieval-time ranking and ultimately to application code — "exactly what Kahl's Cond 3 rules out: external prescription rather than self-maintaining arbitration."

MemMachine (Wang et al., 2026a) — (X) most-recent-state update

The system updates contradicted profile data to the most recent state; no contradiction record, no cascade to episodic records built on the old fact.

  • paper-verbatim: "When new information contradicts existing profile data, the system can update the profile to reflect the most recent state—supporting the knowledge update capability evaluated in benchmarks such as LongMemEval [7]." (Wang et al., 2026a, sec. 4.7, p. 6) (emphasis added)
  • first-hand-read finding: when new information contradicts a profile fact, the profile is updated to the most recent state — no first-class contradiction node, no cascade from a corrected profile fact to the episodic records generated using the old one.

Pinecone Nexus (Pinecone architecture posts) (Ashutosh & Liberty, 2026) — (X)/(N), and read-only consumer

"Deterministic conflict resolution" appears once, as an undocumented feature list-item operating on the knowledge-compilation side; the consuming agent has no commitment graph at all.

  • paper-verbatim (feature enumeration): "...Typed fields. Per-field citations with confidence levels. Deterministic conflict resolution. Output shaped exactly as the agent specified..." (Ashutosh & Liberty, 2026, sec. Introducing Pinecone Nexus) (emphasis added) — first-hand-read note: this is "the only location in the primary sources where 'deterministic conflict resolution' appears... a list item in a feature enumeration, with no further elaboration. The mechanism is not defined anywhere in the primary sources."
  • first-hand-read finding: "There is no arbitration architecture in Nexus at all from the consuming agent's perspective. The agent queries and receives results; it does not maintain a commitment graph, does not detect when received content contradicts prior commitments."

B.4 No claim structure to cascade through

Letta / MemGPT (GitHub + letta.com docs) (Letta, 2025; letta-ai, n.d.) — (S) text blocks

Memory blocks are strings; edits overwrite in place; no derivation graph.

  • paper-verbatim (letta.com/blog/memory-blocks): "memory block values are just strings." (Letta, 2025) paper-verbatim (GitHub base.py): "The memory_rethink command allows you to completely rewrite the contents of a memory block." (letta-ai, n.d.)
  • first-hand-read finding: "Letta has no derivation graph, no concept of downstream commitments, and no retraction primitive. When core_memory_replace overwrites a foundational claim, nothing downstream is flagged."

Anthropic Claude memory tool (platform.claude.com docs) (Memory Tool, n.d.) — (S) manual coherence

Flat free-form text files; coherence is delegated to the agent's manual effort each session.

  • paper-verbatim: "Note: when editing your memory folder, always try to keep its content up-to-date, coherent and organized. You can rename or delete files that are no longer relevant. Do not create new files unless necessary." (Memory Tool, n.d., sec. Prompting guidance) (emphasis added) — developer-injected instruction to maintain coherence by hand, not a structural mechanism.
  • first-hand-read finding: "There is no mechanism for detecting when a memory file contradicts another memory file. There is no mechanism for identifying which conclusions depended on a retracted premise."

B.5 The aspiration, and the genuine absence

2026 Memory survey — Du (Du, 2026) — (N) named, not solved

A survey, not a system: it names contradiction-handling as something robust systems need, citing no system that does it, and proposes only external fixes.

  • paper-verbatim (§7.3): "Robust systems need temporal versioning (prefer the newest record), source attribution (user statement > agent inference), contradiction detection (flag conflicts for resolution), and periodic consolidation..." (Du, 2026, sec. 7.3, p. 9) (emphasis added) — normative ("need"), not descriptive ("do").
  • paper-verbatim (their §9.3): "Future systems need external validation... uncertainty quantification... adversarial probing... and expiration policies" — the proposed fixes are all external interventions, not structural primitives. The survey names the C3 problem and cites zero solutions.

TierMem (Zhu et al., 2026) — (N) contradiction not a construct, no cascade

TierMem's write-back has a step named Conflict Retrieval, but it only retrieves the most similar existing summary units; an LLM memory manager then SKIPs a fact already entailed, UPDATEs a vague entry with added detail (merged text that inherits provenance), or ADDs a distinct event (the page says the manager outputs "one of four actions" but names only these three). None of the named steps detects or resolves a contradiction, and nothing marks or revisits downstream claims (Zhu et al., 2026, App. A.6, p. 12). First-hand-read finding: "No foundational-retraction-flagging mechanism exists. When a Tier-1 summary [TierMem's own tiering, unrelated to Ch4 §4.5's tiers] is UPDATEd, there is no cascade to dependent summaries or downstream uses of the old content." Verdict: C3 not satisfied — in the steps described, contradiction is absent as a construct; the word 'conflict' names a retrieval step.


B.6 Summary

System C3 failure mode Verbatim available?
Kumiho Advisory / opt-in (revision exists, not automatic) yes (AnalyzeImpact, Depends_On "may be")
SCM No resolution (aspirational flag, trigger-only) yes (§3.2 vs §3.5)
GAM Detect-only (no resolution prose) partial — contradict edge yes; resolution-gap = analysis
Zep/Graphiti Invalidates directly contradicted edge(s); no cascade yes (§2.2.3)
Mem0 v3 External delegation (retrieval ranking) yes — clearest in survey
MemMachine External delegation (most-recent-state update) yes ("most recent state")
Pinecone Nexus External/undocumented; read-only consumer yes (list-item) + analysis
Letta No claim structure (strings) yes ("just strings")
Anthropic memory tool No claim structure (manual coherence) yes ("keep... coherent")
Memory survey 2026 Named as aspiration, no implementation yes (their §7.3, §9.3)
TierMem Contradiction not a construct; no cascade none quoted

Across the surveyed set, the pattern is uniform: the field has the primitives (detection, typed edges, even Kumiho's self-claimed formal AGM compliance) but stops short of automatic arbitration as internal state-change. What separates ENGRAM is operational binding (Ch4 §4.4), and it has three parts, not two: the cascade marks every affected node without being asked; auto-surfacing puts those marks back in front of the agent without an intentional reach; and the cascade machine drives the marked region to adjudication during sleep — walking the frontier in dependency order until no pending marker remains, as a mandatory phase of every consolidation cycle.

The third part is where the contrast with the closest competitor is sharpest, and it is worth stating plainly because the two systems fail and succeed at the same place in the architecture. SCM's sleep cycle is triggered by conflict density and does not resolve conflicts; ENGRAM's sleep cycle is where resolution happens. Both systems reached for the same biological analogy and put conflict handling in the offline phase; only one of them closes the loop there.

One precision, since overclaiming here would repeat the very error this appendix charges to others: ENGRAM's machine does not decide what is true. It finds what needs deciding, orders the decisions so that none rests on a premise still awaiting its own, computes what can be computed, and puts each remaining judgement to the agent. The claim is that no affected node stays quietly unexamined — not that arbitration happens without a mind.