Cleanup Memory

Two surfaces:

• External — ~/.claude/projects/<slug>/memory/ (per-machine; never hardcode the path)
• Repo — memory/, CLAUDE.md, docs/architecture/design-rules.md, docs/architecture/code-review-rules.md, docs/sections/

Repo memory/ is the source of truth for durable project rules (syncs via git). External holds only about-Peter / user-pref entries and active-PR working state. Read memory/META.md before working.

Arguments

• (none) — both phases: external → repo → consolidated report → ask → execute
• external — Phase 1 only
• repo — Phase 2 only
• report-only — full report, no changes

Hard Rules

• Never hardcode the external memory path. Derive at runtime (see below).
• Never delete a file before confirming the rule survives somewhere. Diff content; don't match by filename alone.
• Always work in the workspace memory/process/always-use-worktree.md selects. Locally that is a worktree at .claude/worktrees/cleanup-memory-<date> and the main checkout is never edited; in a cloud run it is the repo root on the branch.
• Migration PRs never modify the same file twice. Add atom (PR A), merge, then consolidate (PR B).

Discovering the External Memory Directory

Slug = absolute git root with :, /, \ replaced by -. Example: H:\source\Humans → H--source-Humans.

PROJECT_ROOT="$(git rev-parse --show-toplevel)"
REPO_NAME="$(basename "$PROJECT_ROOT")"
EXT_DIR=""
for d in "$HOME"/.claude/projects/*/memory; do
  [ -f "$d/MEMORY.md" ] || continue
  parent_slug="$(basename "$(dirname "$d")")"
  if echo "$parent_slug" | grep -qi "$REPO_NAME" \
     || grep -qiE "(humans|nobodies)" "$d/MEMORY.md" 2>/dev/null; then
    EXT_DIR="$d"; break
  fi
done
if [ -z "$EXT_DIR" ]; then
  SLUG="$(printf '%s' "$PROJECT_ROOT" | sed 's|[:/\\]|-|g')"
  candidate="$HOME/.claude/projects/$SLUG/memory"
  [ -d "$candidate" ] && EXT_DIR="$candidate"
fi

If still empty, ask Peter — do not guess. Phase 1 has nothing to do on a fresh machine; skip to Phase 2.

---

Phase 1 — External Memory Scan

Classify every file in $EXT_DIR:

Bucket	Criteria	Action
A. Already in repo	Body substantively present in a repo atom (verify content, not filename)	Delete from external after Peter approves
B. About-Peter / user-pref	Working style, tool preferences, interaction rules	Keep
C. Active-PR / ephemeral	Mentions in-flight branch/PR/worktree or "in progress" state	Keep
D. Durable rule, not in repo	Imperative rule with Why:/How to apply:, architecture/process/code-convention	Migrate

If a repo atom is weaker than the external version (drops a constraint or "why"), treat as D and propose strengthening the atom — don't just mark it A.

Procedure

1. ls "$EXT_DIR". Read MEMORY.md first — it indexes everything and often pre-classifies files.
2. Read every *.md under $EXT_DIR.
3. Cross-reference against memory/**/*.md by content (naming conventions diverge: external uses feedback_foo_bar.md, repo uses bucket/foo-bar.md).
4. Build a per-file table: filename | bucket | reason | proposed action. Surface it to Peter and pause.
5. Apply approved actions:
   • A deletions: rm "$EXT_DIR/<file>". Update $EXT_DIR/MEMORY.md to drop index lines.
   • D migrations: Enter the workspace — locally create the worktree and EnterWorktree into it (mandatory; without this, edits land in the main checkout), in a cloud run just check the branch out in the repo root — then write memory/<bucket>/<kebab-name>.md (frontmatter per memory/META.md), add INDEX line (alphabetical within bucket), commit, push, open PR. Do NOT delete the external file in the same PR — wait until the migration PR merges.

---

Phase 2 — In-Repo Hygiene Scan

Files in Scope

memory/INDEX.md, memory/META.md, memory/**/*.md, CLAUDE.md, docs/architecture/design-rules.md, docs/architecture/code-review-rules.md, docs/architecture/coding-rules.md, docs/sections/*.md

Checks

Each finding: severity (BLOCK / IMPORTANT / NIT), location, what's wrong, proposed fix.

1. Dead INDEX entries. Every memory/INDEX.md line → verify file exists. Missing → BLOCK.
2. Orphan atoms. Every memory/<bucket>/*.md → verify INDEX entry exists. Missing → BLOCK.
3. Description sanity. Each atom needs non-empty description: frontmatter (BLOCK if missing). INDEX one-liner must share at least one substantive noun with the atom's description — catches drift. Do NOT flag wording differences; by design the INDEX line is a tighter compression.
4. Bucket misplacement. Atom lives under wrong bucket (architecture/, code/, process/, product/) per META.md definitions → IMPORTANT.
5. Content duplication. Same-bucket pairs (and likely cross-bucket conflicts) with substantive overlap → IMPORTANT, propose merge or split.
6. Atom vs constitution drift. For each atom, check against docs/architecture/design-rules.md:
   • Consistent → fine
   • Conflicts → BLOCK
   • Strictly redundant (verbatim, no new "why"/"how to apply") → NIT, propose deletion or 1-line pointer
7. CLAUDE.md bloat. Target ~80 lines (per memory/META.md). Over by >20% → IMPORTANT. Candidates to atomize: bulleted rules under "Critical:"/"Important:"/"Rule:" headings, detailed concept sections for narrow tasks. Keep: architecture overview, build commands, terminology, git workflow basics.
8. Stale stub. docs/architecture/coding-rules.md should be a redirect-only stub to memory/INDEX.md. Any crept-in content → IMPORTANT.
9. Stale cross-references. Grep for coding-rules.md references pointing at it for substantive content → IMPORTANT.
10. Atom size. Atoms >80 lines likely belong in design-rules.md → NIT.
11. Section invariants alignment. Each docs/sections/*.md — dead citations to atoms or design-rules sections → IMPORTANT.

Outputs

Consolidated report grouped by severity → present to Peter → apply approved fixes on a new branch + PR, in the same workspace pattern as Phase 1. BLOCK first; IMPORTANT next; NIT only if Peter opts in.

---

End-to-End Pacing (no arguments)

1. Phase 1 inventory + classification → report → pause
2. Execute Phase 1 approved actions (migration PR; deletions queued post-merge)
3. Phase 2 audit → report → pause
4. Execute Phase 2 fixes in a separate worktree + PR (don't stack with Phase 1 — overlapping files)
5. After both PRs land, complete Phase 1 deletions of migrated duplicates

report-only: stop after each phase's report.

What This Skill Is NOT

• Not for inventing new rules — use memory/process/rules-maintenance.md
• Not a CLAUDE.md rewrite — Phase 2 finds candidates; the rewrite is a separate focused PR
• Not a doc-freshness sweep — that's /freshness-sweep