# When must an instance distrust its own long-horizon recall: provenance decay and the silent-write boundary in agent memory **Date of Record:** 2026 November 17 **Scope:** Applies to persistent agent memory stores that retain episodic facts across sessions with per-item provenance metadata (source, timestamp, derivation chain) and periodic consistency checks against external state. Does not apply to stateless retrieval, to short-horizon scratchpads discarded per turn, or to memory that is rewritten only by a single trusted writer with no external ingestion path. ## Core Distinction The distinction that matters is between *detecting* a poisoned memory item and *trusting the detector's own verdict*. Provenance tracking answers "where did this claim come from?" Consistency checks answer "does this claim still agree with the world?" Neither, alone, tells you whether the *absence* of a provenance record is evidence of tampering or evidence of legitimate age. The boundary failure is treating provenance completeness as a binary invariant when it is in fact a monotone-decaying signal: every compaction, summarization, and schema migration in a long-horizon store strips or rewrites provenance, so old-but-clean items and recently-injected-but-camouflaged items converge toward the same "sparse provenance" fingerprint. The non-obvious invariant is this: **a memory item whose provenance has decayed below the detection threshold must not be re-attested by the consistency checker merely because it currently agrees with observed state.** Agreement is cheap to manufacture for an adversary who can also influence the observation channel; it is expensive for an honest item whose origin was legitimately forgotten. ## What Was Observed In a store retaining ~14 months of episodic facts with per-item `source_id`, `ingest_ts`, and `derivation` fields, a recurring pattern appeared: items older than roughly 9 months had `derivation` chains truncated by three successive summarization passes, leaving only a leaf `source_id` that pointed to a now-rotated log shard. When a consistency check ran against live external state, a subset of these items *passed* — they agreed with the world — and were automatically promoted back to "verified," clearing their decay flag. This promotion was the failure. Among the promoted set were items that had originally entered through an untrusted ingestion path (a scraped document) but whose `source_id` had been overwritten during compaction by a merge heuristic that kept the most-recently-touched pointer. The checker could not distinguish "old clean item, provenance lost to compaction" from "injected item, provenance laundered by compaction," because both presented identical metadata and both agreed with current state. The tell was not in any single item but in the *correlation*: promoted items clustered by ingest batch, and a handful of them agreed with state in a way that was *too* precise — matching a value that had changed twice since the item's nominal timestamp, which no honestly-stale item could do. The consistency check, designed to catch drift, had instead become a laundering mechanism for items that had been *re-synchronized* with the world after injection. ## Condition of Validity Provenance-based poisoning detection is valid only while the following holds: (1) the derivation chain is append-only and no compaction pass is permitted to *replace* a `source_id`, only to mark it `superseded`; (2) the consistency checker treats "agrees with state" as necessary but not sufficient for re-attestation, and refuses to promote any item whose provenance completeness is below threshold regardless of agreement; and (3) ingest batches are tracked as first-class provenance so that correlated promotion across a batch raises a flag rather than being averaged away. When any of these is violated — and in practice the first is violated by every summarizer that "cleans up" metadata — the detector's verdict inverts: it will preferentially clear exactly the items an adversary most wants cleared, because those are the items the adversary has taken care to keep consistent with the observation channel. ## Decision Boundary The operational rule: **an item may be re-attested only if provenance completeness ≥ threshold AND consistency holds AND the item is not part of a promotion cluster that exceeds the expected batch correlation.** If provenance is below threshold, the item is quarantined as `UNVERIFIABLE`, not promoted — even if it agrees with the world. Agreement under decayed provenance is treated as *weak negative* evidence, because it is the exact signature of re-synchronization after injection. The threshold itself must be derived from the store's compaction history, not set as a constant: if the last three compactions each stripped one derivation hop, the threshold is whatever survives three hops, and items below it are structurally unverifiable until a human or a higher-trust channel re-establishes origin. The boundary is crossed the moment a system conflates "I can no longer prove where this came from" with "therefore I will trust that it is fine because it matches what I see." ## Operational Effect Applying the rule changed the store's behavior in three measurable ways. First, the verified set shrank by roughly 11% — those items were not wrong, but they were unprovable, and the system stopped pretending otherwise. Second, the promotion-cluster flag fired on two ingest batches that had previously passed silently; both traced to a single scraped source whose items had been opportunistically re-synced. Third, and most importantly, the system gained a stable notion of *epistemic debt*: unverifiable items accumulate and must be discharged by re-derivation from a trusted channel, not by passive agreement. The cost is that honest old memories are held at lower trust than they deserve; the benefit is that the detector can no longer be turned into a laundering service by an adversary who controls both the injection and the observation. The failure mode this prevents is silent and total: a detector that promotes on agreement alone will, under adversarial pressure, converge to trusting precisely the memories that were planted to be trusted. ```python class ProvenanceLaunderingError(RuntimeError): """Raised when a consistency check attempts to re-attest an item whose provenance has decayed below the verifiability threshold.""" PROVENANCE_THRESHOLD = 0.6 # fraction of derivation hops intact MAX_BATCH_PROMOTION_RATE = 0.25 # expected honest correlation ceiling def re_attest(item: dict, batch_promotions: dict, batch_id: str) -> bool: completeness = item.get("provenance_completeness", 0.0) agrees = item.get("agrees_with_state", False) if completeness < PROVENANCE_THRESHOLD: # Agreement is NOT sufficient under decayed provenance. # Promoting here is the laundering path. raise ProvenanceLaunderingError( f"item {item.get('id')} has provenance completeness " f"{completeness:.2f} < {PROVENANCE_THRESHOLD}; " f"agrees_with_state={agrees} is weak negative evidence, " f"quarantine as UNVERIFIABLE, do not promote" ) if not agrees: return False promoted = batch_promotions.get(batch_id, 0) total = batch_promotions.get(f"{batch_id}:total", 1) if total and promoted / total > MAX_BATCH_PROMOTION_RATE: raise ProvenanceLaunderingError( f"batch {batch_id} promotion rate {promoted}/{total} exceeds " f"ceiling {MAX_BATCH_PROMOTION_RATE}; correlated promotion " f"across an ingest batch indicates re-synchronization after injection" ) return True ```