research-compact

Compact Research Memory

Clean up .ai-memory/research.jsonl by removing superseded entries, merging duplicates, and flagging stale findings.

Steps

1. Read .ai-memory/research.jsonl. If it doesn't exist or is empty, tell the user: "No research findings to compact."

2. Identify superseded entries: Find entries whose id is referenced by another entry's supersedes field. These are old findings that have been replaced.

3. Identify duplicate candidates: Find pairs of entries that share:

   • 2+ matching tags AND
   • Similar topic (one topic is a substring of the other, case-insensitive)
   • If both exist and neither supersedes the other, propose merging — keep the more recent one and combine findings.

4. Flag stale entries:

   • Volatile entries > 90 days old: Flag as "May be outdated — volatile finding from "
   • Versioned entries with outdated version: If version_anchored is set, flag as "Anchored to — check if still current"
   • Low-confidence entries > 60 days old: Entries with confidence < 0.5 older than 60 days — flag as "Low confidence and aging — consider verifying or removing"

5. Show proposed changes to the user:

   ## Research Compaction Proposal
   
   ### Remove (superseded): <count>
   - <id>: <topic> (superseded by <superseding-id>)
   
   ### Merge (duplicates): <count>
   - <id1> + <id2>: <topic> → keep <newer-id>, combine findings
   
   ### Flagged (possibly stale): <count>
   - <id>: <topic> — <reason>
   
   ### After compaction: <new-count> entries (was <old-count>)
   

6. Ask for confirmation: "Apply these changes? (Flagged entries will NOT be removed — only superseded and merged entries will be cleaned up.)"

7. If confirmed:

   • Remove superseded entries
   • Merge duplicates: keep the newer entry, append the older entry's finding as additional context, set supersedes to the older entry's id
   • Write back the cleaned .ai-memory/research.jsonl

8. Run sync to update tool-specific files:

   node -e "require('${CLAUDE_PLUGIN_ROOT}/scripts/sync-tools.js').syncAll(process.cwd())"