Audit The Vault

How to audit the MemoryVault for off-spec entries

Note

Goal: Run the read-only vault lint, read the categorized report it writes under _meta/, and apply the suggested fixes by hand — the lint never edits the vault. Prereqs: agentm v4.9.0+ (ships V4 #33), python3 on PATH, and a reachable vault (MEMORY_VAULT_PATH set, or pass --vault PATH). The lint reads only; it surfaces candidate fixes for you to review and apply.

Steps

1. Preview the findings (optional). Run the lint and print findings to your terminal — text for skimming, json for piping:

   python3 harness/skills/memory/scripts/vault_lint.py --format text
   python3 harness/skills/memory/scripts/vault_lint.py --format json

   Narrow the corpus with --scope (all · always-load · projects · personal-private; default all). Point at a specific vault with --vault PATH if MEMORY_VAULT_PATH is unset.

2. Write the audit report. Add --audit to write a grouped operator-review report instead of printing:

   python3 harness/skills/memory/scripts/vault_lint.py --audit

   The report lands at <vault>/_meta/vault-lint-<YYYY-MM-DD>.md (override with --out PATH). The command prints a one-line summary (N error · N warn · N info across N entries → <path>). This file is the only thing the lint writes — it never touches an entry.

3. Read the report. Open <vault>/_meta/vault-lint-<YYYY-MM-DD>.md. Findings are grouped by severity (Errors → Warnings → Info), then by check, then collapsed by identical message so a repeated pattern (e.g. one unknown key across 30 entries) shows once with the affected entry list. Each finding names the entry path and a suggested fix.

4. Apply fixes by hand. Edit the flagged entries yourself. The lint applies nothing — every suggestion is advisory and operator-gated. Re-run step 1 to confirm a clean pass.

For what each finding's check_id and severity mean, see Vault lint checks.

Related

• Vault lint checks reference — the check catalog: id / severity / what each checks / suggested-fix shape.