memory-gc

memory-gc — garbage-collect committed memory

On this project the "user-local" memory store is committed to the repo — it is shared team context, not a private scratchpad. That means it rots like any doc: facts duplicate CLAUDE.md, reference closed issues, describe subsystems that moved, or accrete sprint mechanics that belong in the sprint skill. This skill audits every memory with evidence, then applies only the safe changes.

Canonical store path (symlinked from ~/.claude/projects/<slug>/memory/): <repo>/.claude/memory/ — MEMORY.md is the index (one line per memory); each line links a per-fact *.md file. Some facts live inline in MEMORY.md with no backing file.

Why a workflow (don't do this single-threaded)

The highest-value verdict — "this is OBE, delete it" — is also the easiest to get wrong by guessing. Every memory that references an issue/PR/commit/flag must be verified against live state (gh, git, codebase grep) before it can be called outdated. That is dozens of independent gh calls; fan them out to one sonnet worker per memory so each gathers its own evidence in its own context. The orchestrator only ever sees the conclusions.

Method (4 steps)

1. Inventory + integrity (bidirectional)

cd <repo>/.claude/memory
# files not indexed in MEMORY.md (orphans):
for f in *.md; do [ "$f" = MEMORY.md ] && continue; grep -q "($f)" MEMORY.md || echo "ORPHAN: $f"; done
# index lines whose file is gone (broken links):
grep -oE '\]\([a-z0-9_-]+\.md\)' MEMORY.md | sed -E 's/\]\(|\)//g' | while read l; do [ -f "$l" ] || echo "MISSING: $l"; done

Also enumerate the inline facts (MEMORY.md lines with no link) — group them by section; they get audited too. Enumerate the comparison targets that actually exist: the project CLAUDE.md, any package/dir CLAUDE.md (test/, packages/*/), and the sprint skill (SKILL.md + references/*). Do not compare against a parent repo's CLAUDE.md — that's a different repo's concern.