← ClaudeAtlas

memory-hygienelisted

Use when auditing, pruning, correcting, refreshing, or automatically applying approved evidence-based cleanup to native harness memory.
lawzava/megapowers · ★ 4 · AI & Automation · score 76
Install: claude install-skill lawzava/megapowers
# Memory Hygiene Keep native memory factual, dated, and traceable. This workflow audits and applies approved cleanup. It cannot prove a remembered current-state claim is still true. ## Audit read-only Start read-only. Identify active memory, indexes, and available source records. Do not guess provider paths or treat an index as evidence. Quarantine every new candidate before promotion. Retain only `direct-statement`, `direct-observation`, `source-backed`, or `history-entry-only` evidence. Do not retain `inferred`, `speculative`, `unknown`, or `contested` claims in active memory. A missing transcript permits `history-entry-only` evidence with exact dated history metadata; it does not support reconstructed content. For each retained or revalidation record, capture its origin, source, `observed_at` date, and scope. Mark volatile claims and set a bounded `max_age_days`. Revalidate expired volatile claims against an authoritative source before retention. Memory cannot refresh itself. Conflicts must remain preserved and surfaced. Never resolve silently or overwrite one source with another. Do not put secrets or credentials in the manifest or memory. ## Validate the decision Create a temporary audit manifest using [references/audit-manifest.md](references/audit-manifest.md). Resolve `scripts/memory-audit.go` beside this `SKILL.md`, then run: ```bash go run "$memory_audit_tool" --input "$audit_manifest" --as-of YYYY-MM-DD ``` The validator rejects unsupported evidence, expi