Imported from knofler/api (AI/skills/memory-retrieve/Users/rummanahmed/Dropbox/Dev/PROJECT/CODE/_MY_PROJECT/AI/skills/memory-retrieve/SKILL.md). Install upstream with npx skills add knofler/api --skill memory-retrieve. Copyright stays with the author.

Memory Retrieve Playbook

When to Use

• Starting a task and need prior context, patterns, or preferences
• Looking for past error resolutions to avoid repeated investigation
• Searching for user preferences before making decisions
• Building context for a specialist agent before dispatching work

Prerequisites

• Memory directory structure exists: memory/{patterns,preferences,errors,context}/
• Each subdirectory has a populated index.json file
• Query must include at least one keyword or tag

Playbook

1. Parse Query

• Extract keywords from the query string (lowercase, remove stop words)
• Identify explicit tags if provided (prefixed with tag:)
• Determine memory type filter if specified (prefixed with type:)
• If no type filter, search all subdirectories

2. Scan Indexes

• Load index.json from each target subdirectory
• For each entry, compute a relevance score using three factors:
  • Tag overlap (weight 0.5): Jaccard similarity between query tags and entry tags
  • Confidence (weight 0.3): Entry confidence value (0.0-1.0)
  • Recency (weight 0.2): Decay function on last_used — entries used within 7 days get 1.0, within 30 days get 0.7, within 90 days get 0.3, older get 0.1
• Final score = (tag_overlap * 0.5) + (confidence * 0.3) + (recency * 0.2)

3. Rank and Filter

• Sort all matched entries by score descending
• Apply minimum score threshold of 0.2 — discard entries below
• Select top-N results (default N=5, configurable via limit: prefix)
• Load full entry JSON for each result from memory/{type}/{id}.json

4. Update Usage Metadata

• For each returned entry, increment usage_count by 1
• Update last_used to current ISO timestamp
• Write updated entry back to disk
• Update corresponding index.json entry

5. Format Response

• Return results as structured list with: id, type, tags, content summary (first 200 chars), confidence, score
• Include total matches found vs returned count
• Flag any entries with confidence < 0.3 as low-confidence

Output

• Ranked list of memory entries with relevance scores
• Updated last_used and usage_count on returned entries
• Updated index.json files for accessed subdirectories

Review Checklist

• Query parsed correctly with keywords extracted
• All relevant subdirectories scanned
• Scoring formula applied consistently (tag 0.5, confidence 0.3, recency 0.2)
• Minimum score threshold enforced
• Usage metadata updated on returned entries
• Low-confidence entries flagged in response