A memory layer for LLMs that actually remembers

What it does OpenMemory is a self-hosted cognitive memory engine for LLMs and agents. It stores facts, events, preferences, and skills in a local SQLite or Postgres database, then recalls them with context about when things were true and why they mattered. Python and Node SDKs drop into LangChain, CrewAI, AutoGen, or raw scripts.

The interesting bit Most “memory” for LLMs is just RAG: chunk text, embed, retrieve by cosine similarity. OpenMemory treats memory like memory—episodic, semantic, procedural, emotional, reflective—with a temporal knowledge graph that auto-closes outdated facts (CEO Alice becomes CEO Bob, timeline intact) and a decay engine that forgets adaptively instead of using hard TTLs.

Key highlights

• Multi-sector memory model: facts, events, skills, feelings, insights stored and recalled differently
• Temporal knowledge graph with valid_from/valid_to, point-in-time queries, and auto-evolution
• Composite scoring blends salience, recency, and coactivation—not just embedding distance
• Explainable “waypoint” traces show exactly which memory nodes fed a given recall
• MCP server + VS Code extension: Claude, Cursor, Windsurf can query and reinforce memory natively
• Connectors ingest from GitHub, Notion, Google Drive, OneDrive, web crawlers
• Migration tools from Mem0, Zep, Supermemory

Caveats

• 🚧 Project is actively being rewritten; breaking changes and bugs expected
• Some setup friction: backend server requires cloning, npm install, Docker, or Doppler-managed config
• README notes add/search/delete are async, but shows sync-looking examples without await in places

Verdict Build agents, copilots, or personal knowledge tools that need to actually remember over months? This is worth the rough edges. Just need cheap vector search for a chatbot? Stick with your existing RAG pipeline.