Memory Tidy

MEMORY.md is loaded into every session. It is not a file, it is rent you pay on every turn. Past a fixed size the harness truncates it and says nothing: the tail silently disappears and your agent stops knowing things it "remembers". So this is about retention, not tidiness.

Three tiers, and only the first one costs tokens continuously:

• MEMORY.md — hot dispatcher, loaded every turn. Target 60–100 lines, one pointer per line, - [Title](file.md) — hook, ≤150 chars.
• hub-*.md — warm sub-index, not auto-loaded. A domain of ≥3 related lines lives here.
• MEMORY-archive.md — cold, not auto-loaded, grep-only. Done work + coverage for orphans.

Step 0 — ownership (skipping this is how you corrupt someone's memory)

python3 ~/.claude/scripts/memory_scope.py --explain

~/.claude/projects/ holds three things that look identical: your own indexes, byte-identical copies synced from another machine, and throwaway worktrees. Writing into a synced copy means two writers on one file. Declare ownership in memory_scope.json; never infer it. An undeclared machine exits 3 loudly, because a machine that silently tidies nothing looks exactly like a healthy one.

Step 1 — look before you touch

sh ~/.claude/scripts/memory_tidy.sh --dry-run --force

Prints the detectors' verdict and renders the prompt without changing anything. Read the rendered prompt — that is what the model will act on.

Step 2 — the mechanical part (no model, 0 tokens)

python3 ~/.claude/scripts/memory_orphan_cover.py --dry-run
python3 ~/.claude/scripts/memory_orphan_cover.py

An orphan is not lost data, it is unreachable data — which at recall time is the same thing. Pointers go into the archive, which is not auto-loaded, so coverage costs zero live budget. The hook comes from each note's own description: field: the script transcribes, it does not invent.

Step 3 — fold a domain into a hub

A domain with ≥3 related live lines and no hub gets one. Do it in the same pass; do not ask.

python3 ~/.claude/scripts/memory_fold.py --index <MEMORY.md> --map <mapping.json> --dry-run

Judgement (which slug belongs to which domain) lives in the mapping and is yours. Byte-moving is the script's job. This is deliberate: the documented failure mode of folding is summary-grounding — a model rewrites the hook "while it's in there" and the detail that made the pointer findable evaporates. The script never paraphrases, and it refuses to write at all if any folded slug would lose reachability.

Step 4 — judgement (yours, not the script's)

• Keep LIVE only a standing rule that could fire in a session that never opens the hub.
• Archive closed one-offs, finished projects, superseded tools.
• Before archiving, grep the note for open tails (pending|BLOCKED|TODO|⚠). A genuine unfinished action goes on your task board, not into a grave.
• Unsure whether something is still live? Keep it live. Over-pruning costs more than under-pruning.
• A hook cut mid-sentence ("…before «X") is a defect, not compactness. Repair it from the note's description:.

Step 5 — verify, and only then say done

MEMORY_GUARD_EXCLUDE="$(python3 ~/.claude/scripts/memory_scope.py --exclude)" \
  python3 ~/.claude/scripts/memory_guard.py

Must print GREEN. The model's own summary of its work is not evidence — the guard is the judge.

Gotchas worth knowing before you hit them

• A pin list shared across machines. If the pinned note only exists on one machine, every other machine's "a pinned rule fell out!" alarm can never be cleared — and "restoring" the pointer creates a dead pointer. A pin applies only where the note actually lives.
• Generated files counted as memories. Tooling writes a hints file next to the index; counting it as an orphan creates an eternal chore. An alarm nobody can clear teaches people to ignore alarms.
• A logged-out CLI that exits 0. Headless claude -p without a session printed an error and returned success; routines stamped green and produced nothing. Check claude auth status.
• launchd has no catch-up. A calendar job missed while the machine slept is skipped silently — schedule two slots. The second is free when the detectors say "nothing to do".
• Project dirs starting with - are read as flags by CLI tools: md5 -q ./path, not md5 -q path.

Test

python3 tests/test_memory_tidy.py    # 37 checks, ~1s, no LLM, no network, temp fixture