Memory Search

Memory Search

Purpose / When to Activate

Activate when an agent needs to find relevant memory but does not know the exact file path, domain, or DEC ID.

This skill locates memory — it does not load it. After results are returned, the agent invokes memory-retrieve to load the specific files.

Use cases:

• "Which decisions relate to database connection pooling?" → Mode A (frontmatter: domain=infrastructure, tags contains database)
• "Where did we discuss pool exhaustion?" → Mode B (full-text keyword: "pool exhaustion")
• "What decisions are related to DEC-42?" → Mode C (cross-reference: DEC-42 → related_to + mentions)

Process

Mode A — Frontmatter Search

Search YAML frontmatter fields across decisions/DEC-*.md files.

1. Accept query as field-value pairs: {domain: "infrastructure", level: "operational"} and/or {tags: ["connection-pooling"]} and/or {related_to: ["<DEC-id>"]} and/or {status: "active"}
2. Grep frontmatter blocks (between --- delimiters) of all decisions/DEC-*.md files
3. Match files where ALL specified fields match (AND logic)
4. Extract id, title, and matched field values from each hit
5. Rank by: exact tag match > domain match > level match
6. Return top 10 results

Mode B — Content Search

Full-text keyword search across ALL memory files.

1. Accept query as 1-3 keywords (e.g., "pool exhaustion", "scoring formula")
2. Grep all files under contexts/memory/ for keyword matches
3. For each hit, extract:
   • file_path (relative to contexts/memory/)
   • id from frontmatter (if present)
   • title from first # heading
   • excerpt — the matching line ± 1 line of context (~50 tokens)
4. Rank by: number of keyword matches > file recency (updated_at)
5. Return top 10 results
6. Never return more than ~2,000 tokens total

Mode C — Cross-Reference Search

For a given DEC ID, find all files that reference it + traverse related_to (depth 1).

1. Accept query as a single DEC ID (e.g., DEC-42)
2. Direct mentions: grep all files under contexts/memory/ for the literal string DEC-42
3. Frontmatter related_to: grep related_to: lines in decisions/DEC-*.md for the target ID
4. Outbound relations: read the target file's own related_to field → list those DEC IDs
5. Deduplicate and merge results into a single list
6. For each result, extract: file_path, id, title, relevance (direct_mention | related_to_inbound | related_to_outbound)
7. Rank: related_to_inbound > related_to_outbound > direct_mention
8. Return all results (typically <15 files)

Quality Checks

• Never loads full file content — only returns paths, IDs, titles, and short excerpts
• Total output stays under ~2,000 tokens
• Mode A searches frontmatter only (fast, structured)
• Mode B searches content (slower, broader — use sparingly)
• Mode C is bounded to depth 1 (no recursive graph traversal)
• Results are always ranked by relevance
• Never modifies any file

Outputs

search_results — ranked list of memory file references:

- file_path: decisions/DEC-42.md
  id: DEC-42
  title: "Use connection pooling for database access"
  relevance: related_to_inbound  # or: frontmatter_match | content_match | direct_mention | related_to_outbound
  excerpt: "prevents connection exhaustion under load"  # ~50 tokens max, absent in Mode A

Agent receives this list and decides which files to load via memory-retrieve.

Non-Goals

This skill must NOT:

• Load full file contents (that is memory-retrieve)
• Modify any memory files (that is memory-ingest)
• Make decisions about which results to act on (that is the agent's job)
• Perform recursive graph traversal beyond depth 1
• Search outside contexts/memory/

Memory-search finds. Memory-retrieve loads. The agent decides.