LTM Cache Cleanup Skill

When to Use

Use this skill when:

• The user asks why a memory was not stored correctly
• The user wants to inspect, clean up, or repair the LTM cache
• You need to review non-retryable or failed long-term memory cache entries
• Search results or warnings indicate that cached LTM entries need attention

Cleanup Strategy

1. Start with review_ltm_cache_entries.
2. Focus on status="failed" and retryable=false first.
3. Inspect failure_class, last_error, content, and extraction counts.
4. Decide one of these actions:
   • requeue_ltm_cache_entry if the entry should be retried
   • requeue_ltm_cache_entry with replacement_episode if the text is weak or malformed and should be rebuilt as a structured episode
   • delete_ltm_cache_entry if the entry is obsolete, wrong, or should never be ingested
5. If a broken cache entry should be replaced from scratch instead of retried, delete it and then use create_longterm_memory with a corrected structured episode.

Decision Rules

• low_signal_extraction:
  • Usually rewrite the text into clearer factual statements before requeueing
  • Delete instead if the memory is not worth storing
• unknown_error:
  • Review carefully before requeueing
  • Prefer requeueing only if the text itself still looks valid
• Non-retryable validation failures:
  • Fix the underlying dataspace or payload problem before requeueing
  • Delete the entry if it is no longer useful

Episode Repair Guidance

When rebuilding a replacement episode:

• Put durable declarative facts into replacement_episode.facts
• Use explicit names instead of unresolved pronouns
• Include time only when it materially affects validity or retrieval
• Use replacement_episode.context only for short background
• Prefer one coherent episode over many tiny repairs
• create_longterm_memory works best with small, focused, semantically coherent entries
• Rebuild iteratively when needed instead of packing multiple unrelated repairs into one episode

Long-Term Memory Write Rule

• When you need a fresh durable write, prefer a series of focused corrected episodes over one oversized catch-all repair payload.

Tools

review_ltm_cache_entries

Use this to inspect parked or failed cache entries before taking action.

requeue_ltm_cache_entry

Use this to retry a cache entry. Prefer passing a corrected replacement_episode when the original text is weak or malformed.

delete_ltm_cache_entry

Use this when the cache entry should be removed instead of retried.

create_longterm_memory

Use this after deleting a broken cache entry if it is cleaner to create a fresh corrected structured episode instead of requeueing the old one.

search_longterm_memory

Use this to verify whether the repaired memory later became searchable or to inspect surrounding long-term memory context before cleanup.