smart-memory（智能记忆）

Smart Memory v3.1 Skill

Smart Memory v3.1 is a local transcript-first cognitive memory runtime with revision-aware derivation, pinned context lanes, entity-aware retrieval, and bounded prompt composition.

Core runtime:

• Node adapter: smart-memory/index.js
• Local API: server.py
• System facade: cognitive_memory_system.py
• Canonical store: storage/sqlite_memory_store.py plus transcripts/

Core Capabilities

• transcript-first ingest and per-message transcript logging
• typed long-term memory including preference, identity, and task_state
• evidence-backed revision lifecycle decisions and supersession chains
• explicit core and working memory lanes
• entity-aware retrieval with lightweight relationship hints
• deterministic rebuild from transcript history
• hot-memory compatibility projection for working context
• strict token-bounded prompt composition with trace metadata
• inspection endpoints for transcripts, evidence, history, lanes, and eval runs

OpenClaw Integration

Use the native wrapper package in skills/smart-memory-openclaw/.

Primary exports:

• createSmartMemorySkill(options)
• createOpenClawHooks({ skill, agentIdentity, summarizeWithLLM })

The wrapper remains stable while the backend is now transcript-first under the hood.

Tool Interface

1. memory_search

• purpose: query relevant memory through /retrieve
• supports query, type, limit, min_relevance, and optional conversation_history
• health-checks the backend before execution

1. memory_commit

• purpose: persist important facts, decisions, beliefs, goals, or session summaries
• health-checks the backend before execution
• serializes commits to protect local embedding throughput
• queues failed commits in .memory_retry_queue.json

1. memory_insights

• purpose: surface pending background insights
• health-checks the backend before execution
• calls /insights/pending

API Endpoints

Core endpoints:

• GET /health
• POST /ingest
• POST /retrieve
• POST /compose
• POST /run_background
• GET /memories
• GET /memory/{memory_id}
• GET /insights/pending

Transcript and inspection endpoints:

• POST /transcripts/message
• GET /transcripts/{session_id}
• GET /transcript/message/{message_id}
• GET /memory/{memory_id}/evidence
• POST /revise
• GET /memory/{memory_id}/history
• GET /memory/{memory_id}/active
• GET /memory/{memory_id}/chain
• GET /lanes/{lane_name}
• POST /lanes/{lane_name}/{memory_id}
• DELETE /lanes/{lane_name}/{memory_id}
• POST /rebuild
• POST /rebuild/{session_id}
• GET /eval/suite/{suite_name}
• GET /eval/case/{case_id}

Operating guidance

• query memory before speaking when continuity matters
• do not claim prior context unless retrieval actually supports it
• transcripts are canonical, memories are derived
• treat SQLite as canonical runtime storage
• treat JSON as offline export or backup only
• keep CPU-only PyTorch policy intact

Deprecated

Legacy vector-memory CLI artifacts remain deprecated and should not be revived.