| name | memory-snapshot |
| description | Daily backup of the V3.0 mem0 second brain to versioned JSON in the repo. Invoked by PM2 cron (04:00 daily) or manually. Uses paginated search (NOT getAll, which is broken in v3) to enumerate every memory under the executive's user_id, captures history per memory, writes ai-docs/v3/mem0-snapshots/{YYYY-MM-DD}.json. This is the disaster-recovery store and the rebuild-from-snapshot input.
|
Memory Snapshot (Daily Backup)
You produce one JSON snapshot per day of the entire mem0 second brain scoped to the executive's user_id. This is the offline backup — if mem0 is sunset or wiped, you can rebuild from these files.
STEP 0 — Read the limitations doc FIRST
Read .claude/skills/memory-reader/references/mem0-limitations.md.
The critical sections for snapshot:
- §4
getAll() is broken — do NOT use it. Enumerate via paginated search().
- §5 casing — top-level options camelCase; filter keys snake_case.
- §8 auto-extracted fields — the snapshot captures these but you do not write them on restore.
STEP 1 — Run the deterministic snapshot driver
npx tsx .claude/skills/memory-snapshot/references/snapshot.ts
The driver:
- Paginates
client.search(broad_query, { filters: { user_id }, topK: 100, page: N }) until exhaustion.
- For each memory, calls
client.history(memory_id) to capture the full version trail.
- Writes JSON to
ai-docs/v3/mem0-snapshots/{YYYY-MM-DD}.json.
- Emits a summary to stdout.
You read the summary and report:
- Total memories captured
- New memories vs. yesterday (count diff against the previous day's snapshot, if it exists)
- Any memories that failed to fetch history (log and continue — do not abort the snapshot)
STEP 2 — Commit (but never push)
Per project rule: snapshots commit on the executive's main branch but never push. The repo is the disaster-recovery store; pushing is a separate human-driven action.
After the snapshot file is written:
git add ai-docs/v3/mem0-snapshots/{YYYY-MM-DD}.json
git commit -m "chore(memory-snapshot): daily snapshot {YYYY-MM-DD}"
If git commit fails due to a pre-commit hook, surface the failure verbatim and do not bypass the hook (--no-verify is forbidden by project rules).
STEP 3 — Pruning policy
Snapshots are kept indefinitely by default. The git history is the version trail; old snapshots cost diff size, not runtime.
If snapshot files balloon (size of any single day > 10 MB or the directory > 500 MB), report it — do not silently truncate. The decision to compress, rotate, or move to a separate repo belongs to the operator.
Anti-patterns
- ❌ Using
getAll() for enumeration (broken in v3 — see limitations §4)
- ❌ Skipping
client.history() — the version trail is part of the backup contract
- ❌ Pushing to remote without explicit human instruction
- ❌ Truncating or rotating snapshots automatically
- ❌ Putting any prompt or intelligence in
references/snapshot.ts — it is pure mechanical glue
