memory-reviewlisted

Out-of-band review of agent memory, meant to run weekly - reads the week's session transcripts and the mistakes journal, finds what no single session can see (a mistake repeated across sessions, a correction by the user that memory lacks, contradictions between entries, outdated facts, a broken index), applies the safe fixes, proposes the rest, and empties the journal after logging the analysis. Use on "review memory", "weekly memory pass", "consolidate memory", "what did I keep getting wrong this week", or from a scheduled task.
letsloose501/sqs-agent-memory · ★ 1 · AI & Automation · score 74

Install: claude install-skill letsloose501/sqs-agent-memory

# Weekly memory review Find what is invisible from inside one session: contradictions between entries, outdated facts, repeated mistakes. Report to the user in their language; everything written into memory (entries, index lines, the log) is in English. ## Paths - Memory: `${user_config.memory_dir}`: entries, index `MEMORY.md`, draft `mistakes/`, rules in `claude/rules.md`, constitution `README.md` and `SCOPES.md` (protected by a hook: propose, never rewrite). - Transcripts: `~/.claude/projects/*/*.jsonl`, all projects (memory is shared across them). Do not grep them raw: read the digest from phase 1, and open a transcript only to quote evidence. - Vault: `${user_config.vault_dir}`, the source for checking facts. **Change nothing in it.** ## Phase 1: the week's transcripts Build the digest first: `uv run --no-project "${CLAUDE_PLUGIN_ROOT}/scripts/transcript_digest.py" --exclude-session <this session's id> --out <a temp file>`. It takes a session into the week by the **last timestamp inside the file**, not by mtime (a bulk file operation gives old sessions a fresh mtime), skips generated `*evals*` projects, keeps only the user's own messages (system reminders and skill bodies dropped), groups failed tool calls by signature across sessions with `REPEAT` at two or more distinct sessions, and lists tool calls the user rejected. Why not grep: the rules text rides in every turn, so symptom words match every transcript and the search drowns in noise. The digest is a pre-