Real long-term memory for AI agents with cognitive-science-inspired five-sector memory model and temporal knowledge graphs, featuring memory decay, reinforcement, and explainable retrieval.
OpenMemory is a persistent long-term memory system designed for LLMs and autonomous agents. Unlike traditional RAG or vector databases, it employs a cognitive-science-inspired five-sector memory model (Episodic, Semantic, Procedural, Emotional, Reflective), combined with temporal knowledge graphs, memory decay and reinforcement engines, Waypoint association graphs, and other mechanisms to provide a solution closer to human memory characteristics.
Core Features:
- Five-Sector Memory Model: Cognitive science-based hierarchical sector decomposition, not simple vector storage
- Temporal Knowledge Graphs:
valid_from/valid_totime windows, supporting fact evolution and point-in-time queries - Memory Decay & Reinforcement Engine: Sector-aware decay curves, high-signal event triggered pulse reinforcement, attribution tracking
- Waypoint Association Graphs: Unidirectional strongest links (cosine ≥ 0.75), 1-hop graph traversal, weight decay and auto-pruning
- Explainable Retrieval: Composite scoring
0.6×similarity + 0.2×salience + 0.1×recency + 0.1×Waypoint, with recall path tracing - Multi-modal Ingestion: PDF, DOCX, TXT, MD, HTML, audio/video (Whisper API)
- Data Source Connectors: GitHub, Notion, Google Drive, OneDrive, Web Crawler
- Multiple Embedding Providers: OpenAI, Gemini, AWS, Ollama, local models, synthetic fallback
Architecture: Hierarchical Memory Decomposition (HMD v2) five-layer architecture, SQLite storage, supports horizontal sharding and vertical optimization (WAL mode, SIMD vector computation).
Performance: Add 80-120ms, Query single-sector 110-130ms, 100k memories ~500MB.
Deployment: Python SDK (openmemory-py), Node SDK (openmemory-js), Docker Compose, CLI tool (opm).
MCP Integration: Supports Claude Desktop, Cursor, Windsurf, VS Code with tools like openmemory_query, openmemory_store, openmemory_list.