Gotcontext Memory Claim Lifecycle

gotcontext-memory — claim lifecycle

Companion to gotcontext-memory-hitl-honesty and gotcontext-memory-streaming-digests.

Core ideas

• Promotion is HITL — dream only stages; humans accept/reject.
• Rejected claims have claimKey identity — claimKey(targetPath, body) is stable across runs and independent of base_hash so rejects still suppress the next identical body.
• Accepted prefs are settled by path presence — if storeHashes already has targetPath, skip re-create (protects human edits). Do not pretend accepts are archived under claimKey.
• Prevalence without a window is noise — unbounded k/n (e.g. 16/17263) is true and useless.

Rules

1. No resurrection (two mechanisms) — (a) loadSuppressedClaims from proposals/rejected/** via claimKey; (b) preference extract skips when the target file already exists in the store. Red tests in test/dream-suppression.test.ts.
2. claimKey ≠ proposalId — proposal id may include base_hash / evidence material; rejected-claim suppression must not, or CAS drift re-stages the same advice forever.
3. Window prevalence — CLI --max-sessions (default 400) per source via stratified selectDigests (≈2/3 newest + older strata); report k/n against that window only.
4. Sort by evidence strength — before maxProposals slice: evidence.length desc, then path/id. Never order survivors by sha256 id prefix.
5. Preference FP hose — anchors please remember | from now on only; deny pong/ping//health (L11). Bare always/prefer is forbidden without new FP tests.
6. Alive ≠ useful (L7) — receipts must include staged content quality + accepts, not only “dream run succeeded.”

Signals today (regex / counts — not LLM)

Authority

• Lessons L7, L11, L16, L17
• Code: claimKey / runDreamFromDigests in src/dream/run.ts
• docs/features/dream.md, docs/HONESTY.md
• Prior art: selective promotion / late filtering — docs/research/2026-08-09-agent-memory-prior-art.md

Verify

cd gotcontext-memory && npm test -- test/dream-suppression.test.ts test/dream.test.ts