False-invalidation safeguards

Automatic conflict resolution has an asymmetric failure mode. A false duplicate verdict sits visibly in the review queue where a human can veto it. A false invalidation silently removes a memory from recall — the knowledge effectively disappears, and nobody is told. zero-memory treats this asymmetry as a first-class design constraint: a set of safeguards makes the dangerous direction strictly harder to trigger, and a restore path exists for anything that slips through.

Why it exists

The risk is concrete, not hypothetical: an LLM judge comparing a broad summary against a narrow, precise gotcha can plausibly rule that the broad record "supersedes" the narrow one — even though the broad record contains none of the narrow one's actionable detail (exact diagnostic steps, the difference between two environment switches, the command that fixes the problem). Acting on that verdict hides the specific record and produces a net loss of information, which undermines the product's core promise that memory is a reliable first source. The safeguards below exist to make that outcome structurally impossible for the knowledge that matters most.

The safeguards

Protected kinds are never auto-invalidated. A memory of kind gotcha, convention, decision, or preference with an authoritative provenance rank (rank 1 or higher — written deliberately by an agent or a human, not provisionally captured by the watcher) can never lose an automatic resolution. Whatever the judge's verdict — duplicate or supersedes — the pair goes to the review queue. A false duplicate in the queue is visible; a silent invalidation is not.

Rank-based auto-supersede only demotes provisional records. Authority short-circuits the judge only when the losing side is rank 0 (provisional watcher capture). Pairs where both sides carry real authority — a human-ranked record against an agent-ranked one — always go through the judge's verdict plus the protected-kind guard, and when such a pair is a duplicate, the more authoritative side wins, not merely the newer one. The one lane that acts on any relation at any confidence is the both-provisional collapse: two rank-0 captures of the same thing merge with the newest winning — safe by construction, because no authoritative knowledge is involved on either side.

Invalidation demands more confidence than deduplication. Auto-superseding requires judge confidence of at least 0.95, while auto-forgetting an exact-duplicate pair keeps the standard 0.9 threshold. Hiding a record whose content differs from the survivor's is a more destructive act than collapsing two records that say the same thing, so it must clear a higher bar.

The judge distinguishes "replaces" from "complements". The judge's prompt carries an explicit coverage clause: a supersedes verdict is only valid if the replacement covers all of the actionable specifics of the record it replaces — diagnostic steps, commands, configuration differences. A broad summary and a specific diagnosis of the same topic is complementary, not supersedes. High-confidence complementary pairs get a relates_to link between them and both records stay live; nothing is hidden.

Restoring an invalidated memory

If a memory was invalidated — by an automatic action, a review decision, or your own earlier forget — the restore_memory MCP tool brings it back:

• clears the invalidation and supersession marks, so the memory reappears in recall and briefings;
• removes the supersede links pointing at it;
• writes an audit entry recording the restore.

Restore works even for memories that were hidden inline at write time and therefore never had a review-queue entry — no dispute record is required. Records that are merely disputed (a pending conflict in the queue) remain visible in recall with a dispute flag and are resolved through list_conflicts / resolve_conflict instead — for example with a keep_both resolution.

Challenging a memory that misled you

The safeguards above protect knowledge from the judge; the challenge MCP tool covers the opposite direction — a memory the safeguards keep alive but that reality contradicts. The moment a recalled memory proves wrong against the world (a stored fix that fails, a decision the code contradicts), an agent calls challenge(memory_id, reason) — used when the corrected fact is not yet known; when it is, writing it with a supersedes link is the right move instead.

A challenge raises a single-subject dispute (verdict challenged) in the review queue: one memory, no counterpart. The challenged memory stays live and visible in recall — it only ranks lower until adjudicated, because recorded doubt beats silently ignoring a memory. Challenging is idempotent while the dispute is open. The dispute is settled with resolve_conflict: uphold the memory (winner = its id, or keep_both), or retire: true — a reversible retirement with no successor, undone by restore_memory exactly like any other invalidation. Nothing ever retires a challenged memory automatically.

Hygiene raises the same dispute shape on its own evidence: a memory whose misled signals (explicit challenges plus judge-attributed misled verdicts) cross a threshold is queued as a stale suspect, and an external re-check that finds a world fact outdated files one too (see world-fact freshness). In every case the flag is a question for review, never an invalidation. A dismissed suspicion is only re-raised by evidence newer than the dismissal. The dashboard's review page shows pair conflicts; single-subject disputes are triaged through the conflict tools.

How to use it

• restore_memory — soft-recover a hidden memory by id.
• challenge — dispute one of your memories as wrong, stale, or misleading.
• list_conflicts, get_conflict, resolve_conflict — inspect and settle pending disputes; keep_both is the right resolution when both sides are true, retire: true reversibly retires a single-subject dispute's memory.
• The dashboard review queue is the visual surface for pair decisions — see the hygiene overview.

Design notes

• Two invalidation paths, one dangerous. A record can be contested via the review queue (visible, flagged in recall, reversible through resolve_conflict) or hidden directly by a confident automatic verdict (invisible). The safeguards deliberately push every ambiguous case onto the first path; restore_memory exists for the second.
• Thresholds are asymmetric on purpose. The cost of a false positive is not symmetric across actions, so the confidence bars are not either. The real-world incident that motivated this work would have been stopped by the 0.95 threshold alone — and by the protected-kind guard independently.
• Restored memories are respected downstream. The queue-throughput machinery treats a previously restored memory as settled: triage keeps both sides, and already-adjudicated pairs are never re-judged by a later sweep.