Trw Memory Optimize

Sizing Guidelines

The optimal learning count scales with project complexity — do NOT use a fixed target.

Formula: Target = (distinct domain count) × 3-5 entries per domain, with a floor of 20.

How to calculate:

1. Identify distinct topic clusters (e.g., "hallucination", "testing", "ollama", "transcription")
2. Each cluster should consolidate to 3-5 entries depending on depth:
   • Simple domain (few gotchas): 2-3 entries
   • Complex domain (many patterns, edge cases): 5-8 entries
3. A project with 12 domains should target ~50-70 active entries, not 30

Consolidation depth limit: Never merge more than 10-15 entries into a single compendium. If a topic has 60+ entries, create 5-8 sub-topic compendiums (e.g., "hallucination-grounding", "hallucination-detection", "hallucination-mitigation") rather than one mega-entry.

Domain coverage rule: Every distinct domain MUST retain at least 1 detailed entry after optimization. If consolidation would leave a domain with 0 entries, it's too aggressive.

Constraints

• NEVER delete learning YAML files — mark as obsolete status instead
• ALWAYS present the plan and get user confirmation before any changes
• ALWAYS preserve high-impact (>= 0.7) entries unless clearly outdated
• ALWAYS run delivery sync after changes to keep AGENTS.md current
• NEVER collapse all entries in a domain into a single compendium — maintain sub-topic granularity

Assertion Verification Wave (PRD-CORE-086)

After standard pruning and consolidation, run an assertion verification wave:

1. Collect: Identify all learnings with non-empty assertions via trw_recall(query="*", max_results=0) and filter for entries with assertion_status
2. Verify: For each entry with failing assertions, spawn a subagent to investigate:
   • Read the referenced files in the codebase
   • Determine root cause: Is the learning outdated? Is the assertion pattern wrong? Is the code violating the convention?
   • Recommend one of:
     • UPDATE_LEARNING: Learning text needs revision (provide new text)
     • UPDATE_ASSERTION: Assertion pattern is wrong (provide corrected pattern)
     • RETIRE_LEARNING: Knowledge is obsolete (provide reason)
     • CODE_VIOLATION: Code is wrong, learning is right (flag for human review)
3. Apply: For each recommendation, use trw_learn_update() to apply changes (with user confirmation for retirements)
4. Report: Summary table of verification results — passing, fixed, retired, flagged

Notes

• Run after major milestones (strip-downs, sprint completions) when many entries may be stale
• Use /trw-memory-audit first for a read-only preview