memory-audit

mem0 Memory Audit

Analyse the current state of the mem0 memory database and produce an improvement report. Called as /mem0:audit or invoked automatically when the user asks about memory quality.

What to do

1. Retrieve all memories for the current user:

   from mem0 import MemoryClient
   import os
   client = MemoryClient(api_key=os.environ["MEM0_API_KEY"])
   user_id = os.environ.get("MEM0_USER_ID", "user")
   agent_id = os.environ.get("MEM0_AGENT_ID", "")
   
   user_mems = client.get_all(filters={"user_id": user_id})
   proj_mems = client.get_all(filters={"user_id": user_id, "agent_id": agent_id}) if agent_id else {"results": []}
   

2. Evaluate quality across these dimensions:

   Atomicity — Is each memory a single, self-contained fact? Flag: memories with conjunctions ("and", "also", "but") that could be split.

   Specificity — Are memories concrete enough to be useful? Flag: vague entries like "we discussed authentication" or "there was an issue".

   Duplicates — Are there semantically redundant memories? Flag: pairs with >80% semantic overlap.

   Staleness — Do any memories reference things likely outdated? Flag: version numbers, dates, or "current" references.

   Extraction gaps — What important patterns appear in conversation that are NOT in memory? (Review last 5 conversation turns if available.)

3. Produce a structured report:

   ## Memory Audit Report
   
   ### Summary
   - Total memories: N user-scope, M project-scope
   - Quality score: X/10
   
   ### Issues Found
   [List each issue with: memory text, issue type, suggested fix]
   
   ### Extraction Prompt Assessment
   [Is the current extraction prompt capturing the right things?
    What category of facts is missing or overrepresented?]
   
   ### Recommended Changes
   1. Split: "[memory text]" → "[fact 1]" + "[fact 2]"
   2. Delete: "[memory text]" — reason
   3. Update extraction prompt: add/remove/change rule X
   

4. Ask the user if they want to apply the suggested changes. If yes, update memories via client.update() or client.delete().

Notes

• Do not delete without user confirmation.
• The extraction prompt lives in hooks/stop.py as EXTRACTION_PROMPT. If changes are recommended, show the exact diff.
• Run this periodically (e.g. after 10+ sessions) for best results.