AgentSkillsCN

Memory Recall

记忆回溯

SKILL.md

/recall — Memory Search

Search across all structured memory files. Default backend: BM25 scoring with Porter stemming and domain-aware query expansion. Optional: graph-based cross-reference boosting (--graph). Optional: vector/embedding backend (configure in mind-mem.json). Returns ranked results with block ID, type, score, excerpt, and file path.

When to Use

  • Before making decisions (check if a similar decision already exists)
  • When asked about past events, decisions, or tasks
  • To find related context for a current problem
  • To check what's known about a person, project, or tool
  • To explore connections between decisions and tasks (--graph)

How to Run

Basic Search

bash
python3 maintenance/recall.py --query "authentication" --workspace "${MIND_MEM_WORKSPACE:-.}"

Graph-Boosted Search (cross-reference neighbor discovery)

bash
python3 maintenance/recall.py --query "database" --graph --workspace "${MIND_MEM_WORKSPACE:-.}"

JSON Output (for programmatic use)

bash
python3 maintenance/recall.py --query "auth" --workspace "${MIND_MEM_WORKSPACE:-.}" --json --limit 5

Active Items Only

bash
python3 maintenance/recall.py --query "deadline" --workspace "${MIND_MEM_WORKSPACE:-.}" --active-only

What It Searches

  • decisions/DECISIONS.md — All decisions
  • tasks/TASKS.md — All tasks
  • entities/projects.md — Projects
  • entities/people.md — People
  • entities/tools.md — Tools
  • entities/incidents.md — Incidents
  • intelligence/CONTRADICTIONS.md — Known contradictions
  • intelligence/DRIFT.md — Drift detections
  • intelligence/SIGNALS.md — Captured signals

Scoring

Results are ranked by BM25 relevance (k1=1.2, b=0.75) with:

  • Stemming — "queries" matches "query", "deployed" matches "deployment"
  • Query expansion — "auth" expands to include "authentication", "login", "oauth", "jwt"
  • Recency — Recent items score higher
  • Active status — Active items get 1.2x boost
  • Priority — P0/P1 items get 1.1x boost
  • Graph neighbors — With --graph, blocks connected via cross-references to keyword matches get a 0.3x boost (tagged [graph] in output)