Knowledge Audit

Knowledge Audit

Systematic review of stored memories to maintain quality. Finds duplicates, detects contradictions, flags stale entries, and consolidates overlapping memories.

Steps

1. Scope the audit:

   • If topic specified: memory(action="search", query="[topic]") to find all related memories
   • If "all": memory(action="list") to get full inventory, then memory(action="stats") for overview
   • Group memories by tag/category for systematic review

2. Identify duplicates:

   • Search for memories with similar content or overlapping keywords
   • Compare pairs that cover the same topic
   • Decision: keep the more detailed/recent one, delete the other
   • Use memory(action="delete", id="[duplicate-id]") for removals

3. Detect contradictions:

   • Look for memories that make opposing claims about the same topic
   • Examples: "Use library A" vs "Switched to library B", conflicting conventions
   • Decision tree:
     • Both have dates -> keep the newer one (it supersedes)
     • Neither has date -> ask user which is current
     • Both are valid (context-dependent) -> update both to clarify their scope
   • Update the surviving memory to note it supersedes the old one

4. Flag stale entries:

   • Memories referencing specific versions that are now outdated
   • Memories about temporary workarounds that may have been resolved
   • Memories about tools/libraries that have been replaced
   • Action: mark as stale with memory(action="update", ...) adding [STALE] prefix, or delete if clearly obsolete

5. Consolidate overlapping memories:

   • Multiple memories about the same topic that each have partial info
   • Merge into a single comprehensive memory
   • Steps: create new consolidated memory -> verify retrieval -> delete originals
   • Use memory(action="consolidate", ...) if available, otherwise manual merge

6. Produce audit report:

   ## Knowledge Audit — [topic/all] — [date]
   
   ### Summary
   - Total memories reviewed: [N]
   - Duplicates removed: [N]
   - Contradictions resolved: [N]
   - Stale entries flagged/removed: [N]
   - Memories consolidated: [N merged into M]
   
   ### Actions Taken
   - [list of specific changes]
   
   ### Recommendations
   - [any patterns noticed, e.g., "many memories lack WHY context"]
   

Staleness Indicators

• References to specific version numbers (check if still current)
• Contains "temporary", "workaround", "until X is fixed"
• References removed/renamed files, deprecated APIs, old URLs
• Predates a major migration or refactor (check project history)
• Contains "TODO" or "will be" — was it done?

Contradiction Resolution Rules

• Explicit supersession: If memory B says "switched from A to B", delete memory about using A
• Scope difference: "Use X for backend" and "Use Y for frontend" are NOT contradictions
• Evolution: "Started with X" and "Migrated to Y" — keep Y, delete X (unless X context is still relevant)
• Ambiguous: When unclear, DO NOT delete — ask the user

When to Use

• Periodically (monthly or after major project changes)
• When memory search returns confusing or contradictory results
• After a major migration or architecture change
• When starting a new phase of a project
• When memory count grows large and retrieval quality degrades

Phase 3: Temporal KG Audit (v2.0+)

Phase 3 ships a temporal knowledge graph (entities + edges + bitemporal versioning + supersession). The audit checklist now also covers KG hygiene:

KG-aware audit dimensions

7. Stale entities: entity rows that no longer link to any active memory (their last memory_entity_links row points at an archived / superseded memory).

   • Detect: memory(action="entity_search", name="<entity>") returning 0 currently-valid hits → candidate stale.
   • Action: confirm with the user, then delete the orphaned entity row (cascade removes edges).

8. Orphan edges: edges in memory_edges whose source or target memory_entities row was deleted but the edge survived.

   • Detect: server logs show "edge with missing endpoint" warnings, or entity_graph returns nodes referenced from edges that aren't in the nodes list.
   • Action: ask user; either re-extract the originating capture or hand-delete the edge by id.

9. Contradicting / superseded chains: a memory in a supersession chain still surfaces in default memory.get results because someone left valid_to = NULL on an old fact.

   • Detect: memory(action="history", entity_id=<x>) returning multiple rows with valid_to = NULL for the same entity.
   • Action: pick the most-recent / most-correct row and update valid_to on the others to the supersession timestamp.

10. Bitemporal drift: capture has no valid_from set, falls outside the bitemporal index. Usually a pre-Phase-3 row that was missed by backfill, or a manual db.add that bypassed the capture pipeline.

    • Detect: SQL spot check SELECT COUNT(*) FROM memories WHERE valid_from IS NULL.
    • Action: re-run the Phase 3 backfill (MemoryDB._backfill_phase3_temporal) or update the rows manually.

11. Audit-trail integrity: every mutation (insert / update / supersede / delete) should write a memory_audit row with prev_state_hash and new_state_hash. A gap in the chain (audit row absent for an update) signals a bug or out-of-band write.

    • Detect: SELECT m.id FROM memories m LEFT JOIN memory_audit a ON a.memory_id = m.id WHERE a.id IS NULL.
    • Action: investigate which path wrote the row; do NOT rewrite the audit history (preserves tamper-detection guarantees).

When to run the temporal sub-audit

• After running the Phase 1/2 "Knowledge Audit" steps above.
• After importing a Phase 2 passport bundle (legacy schema → can leave bitemporal columns NULL).
• After any large memory(action="capture", auto=True) batch with KG_AUTO_ENABLED=true (entity resolution may need tuning).
• Before exporting a Phase 3 passport bundle (so receiver gets a clean KG).