Review per-project memory monthly, prune stale entries

Auto-saved memory files (~/.claude/projects/*/memory/MEMORY.md) get stale — 'currently working on X' after X shipped. Monthly review + prune.

Memory review monthly

The rule

Once a month, review every ~/.claude/projects/*/memory/MEMORY.md file + any per-project memory notes. Delete stale entries. Not archive — delete, per knowledge-deletion-not-supersession.

Why this exists

Auto-saved memory accumulates false facts over time:

• “Currently working on X” — X shipped 3 weeks ago
• “User prefers Y” — user changed to Z last month
• “Project uses framework A” — migrated to B in the interim
• “Blocked on Z” — Z was resolved in a different session

Claude re-reads memory at session start and treats stale entries as current fact. Result: confident-wrong responses.

Review protocol

Once a month (calendar reminder or on the first commit of the month):

1. ls ~/.claude/projects/*/memory/ — list per-project memory dirs
2. For each MEMORY.md:
   • Read every one-liner pointer
   • For each pointer, ask: “Is this still true?”
   • Delete stale, keep live
3. For each memory file linked from MEMORY.md:
   • Open, check metadata.type
   • feedback/project: verify still applies. Delete if user preference changed OR project completed.
   • user: verify user still matches. Rarely deleted; occasionally updated.
   • reference: verify URL/target still exists. Delete if broken.
4. Commit: chore(memory): prune stale entries YYYY-MM

What COUNTS as stale

• Timeframe: any “currently”/“in progress”/“waiting on” entry older than 6 weeks
• Superseded: entries where a newer decision file in knowledge/ contradicts
• Duplicates: same fact stored in knowledge/ + memory → keep knowledge, drop memory
• Wrong: entries that were factually incorrect when written (learned in a later session)

What does NOT count as stale

• Genuine long-term preferences (user role, permanent taste rules)
• Reference URLs still live
• Feedback captured with “Why” + “How to apply” that still applies

Anti-patterns

• ❌ Never reviewing (default) — memory becomes a fiction
• ❌ Reviewing but not deleting — accumulating “just in case”
• ❌ Moving stale memory to knowledge/ “in case it’s useful” — if it’s not durable, delete
• ❌ Auto-review script that deletes anything older than N days without human eyes — misses genuine long-term facts

Automation aids (allowed)

• Script that surfaces memory files older than 30 days for review
• Script that flags memory entries containing “currently”/“in progress”/“today”/“this week”
• Never a script that AUTO-DELETES

Cross-refs

• self-update-rule — the write-side complement
• knowledge-deletion-not-supersession — deletion pattern
• own-memory-rent-intelligence — the reason memory hygiene matters
• Source: Nate Herk YouTube 2026-07-02 — his “kept thinking I was still working on videos I’d already posted” observation

Edit on GitHub · Back to index