Memory Quality Auditor
Audit the memory system as a unified retrieval layer (STM/MTM/LTM files + index + spawn citation outcomes).
Scope
- Retrieval drift signals
- stale memory ratio
- evidence injection coverage
- citation usage/groundedness continuity
Workflow
- Read memory artifacts and latest eval reports.
- Compute quality metrics and threshold status.
- Emit remediation backlog with TDD checks.
- Record findings in memory and optional evolution recommendation.
Iron Laws
- ALWAYS establish a baseline metric snapshot before auditing — drift is only meaningful relative to a prior measurement; auditing without a baseline produces absolute numbers that cannot identify regression.
- NEVER close a memory finding without re-running the affected retrieval query — closing without verification creates false improvement metrics and masks persistent degradation.
- ALWAYS include citation-groundedness checks in every audit run — uncited memory injections are the primary source of hallucination in agent spawns; skipping this check leaves the highest-risk failure mode undetected.
- NEVER audit only the STM tier — degradation often originates in MTM/LTM promotion corruption; all three tiers must be sampled in every full audit cycle.
- ALWAYS emit TDD-ready remediation items with a failing-test condition and expected metric threshold — vague findings ("memory quality is low") cannot be actioned by any agent.
Anti-Patterns
| Anti-Pattern | Why It Fails | Correct Approach |
|---|---|---|
| Auditing without a baseline | Cannot distinguish regression from steady-state; all findings are ambiguous | Snapshot current metrics at session start; compute delta against the previous run |
| Closing findings without re-check | Produces false-positive resolution; degradation persists silently behind green metrics | Re-run the specific retrieval query after each remediation; close only on confirmed green metric |
| Skipping citation groundedness | Citation failures are the leading cause of agent hallucination; missing this check omits the highest-severity defect class | Include citation_coverage and grounded_ratio metrics in every audit report |
| Full-mode audit on every spawn | Full audit is expensive; running it unconditionally inflates cost and slows workflows | Use --mode summary for routine checks; reserve --mode full for scheduled or triggered audits |
| Auditing STM only | MTM/LTM corruption is invisible in STM-only scans; stale LTM entries contaminate future sessions | Sample all three tiers: STM (current session), MTM (last 10 sessions), LTM (permanent summaries) |
Memory Protocol (MANDATORY)
Before starting: Read .claude/context/memory/learnings.md
After completing:
- New pattern →
.claude/context/memory/learnings.md - Issue found →
.claude/context/memory/issues.md - Decision made →
.claude/context/memory/decisions.md
ASSUME INTERRUPTION: If it's not in memory, it didn't happen.