Cognee

Cognee gives AI agents memory across sessions so they can reuse project context, decisions, and lessons from earlier work. It's for developers building agents and teams whose knowledge sits across documents, code, and conversations.

• Connected retrieval: Search related facts and relevant passages using graph, vector, and relational retrieval, with code memory that captures symbols and dependencies.
• Agent connections: Claude Code and Codex plugins retain context across runs; Cursor, OpenClaw, LangGraph, and compatible MCP clients can use the same memory.
• Shared team knowledge: Bring Slack, GitHub, and Linear content into memory that agents can search alongside other material.
• Domain rules and learning: Custom data models and ontologies describe your domain, while session distillation preserves accepted lessons and feedback adjusts later recall.
• Answers and deletion: Retrieve supporting passages or generate answers with an optional local or hosted LLM, and remove individual items or datasets.

The free Python package is open source under Apache 2.0. Local memory processing uses GLiNER and a local embedding model on your CPU without an API key; both models download on first use. Choosing OpenAI models and embeddings sends processing and answer requests to OpenAI.

You can run Cognee on your own infrastructure, including Docker and on-premises deployments. Cognee Cloud is a separate managed hosting service.