insights

Insights

Mine memsearch history for recurring patterns and promote valuable ones into auto-memory. If no patterns are found, say so. Do not invent patterns.

What to look for

Search memsearch for patterns that repeat across sessions:

• Corrections that appear 3+ times become feedback memories
• Preferences stated or demonstrated repeatedly become user memories
• Ongoing goals and constraints become project memories
• External system pointers become reference memories

How memsearch search works

Memsearch has three layers of progressively deeper recall:

L1: Search

Run memsearch search "<query>" --top-k 10 --json-output to find relevant memory chunks. Start with broad queries:

• "CORRECTION:" or "user corrected" or "user wanted"
• "PREFERENCE:" or "user prefers" or "always"
• "DECISION:" or "going forward" or "agreed"
• "BLOCKER:" or "do not" or "doesn't work"

L2: Expand

For promising results, run memsearch expand <chunk_hash> to see the full markdown section with context.

L3: Transcript drill-down

If you need exact wording:

• memsearch transcript <jsonl_path> to list all turns
• memsearch transcript <jsonl_path> --turn <uuid> --context 3 for a specific turn

Process

1. Search (L1): Run broad queries across the project collection to find candidate patterns. Look for themes that appear in multiple results.

2. Expand (L2): For each candidate, expand the chunk to see full context. Verify the pattern is real, not a one-off.

3. Drill down (L3, optional): If the pattern involves a specific correction or decision, drill into the transcript for exact wording and reasoning.

4. Check existing memory: Read the current auto-memory files to know what's already captured. Don't duplicate.

5. Promote: For each new pattern worth keeping, save it using the auto-memory system following the instructions in your system prompt. Use the appropriate type:

   • feedback for corrections, preferences, and style guidance
   • project for ongoing initiatives, goals, and constraints
   • user for role, expertise, and working style
   • reference for pointers to external systems and resources

6. Prune: Check existing memories against current memsearch data. If a memory claims something that contradicts recent sessions, update or remove it.

Judgment calls

• Only promote patterns that appear 3+ times or that the user explicitly emphasized. One-off corrections are noise.
• Prefer why over what. "Don't use em dashes" is less useful than "Don't use em dashes. Why: user finds them visually noisy."
• If a pattern is already in CLAUDE.md or derivable from code, skip it.
• If unsure, don't save it. Under-saving is better than cluttering memory.

Output

Report what you found and what you promoted. Be brief.