Cognee
Open-source AI memory engine that turns documents and app data into a knowledge graph plus vectors for agents
Overview
Cognee is an open-source memory layer for AI agents. You feed it documents, conversations, code or data from tools like Slack, Notion, Linear and GitHub, and it extracts entities and relationships into a knowledge graph while also storing embeddings. Agents then query that memory with graph-aware retrieval, which handles multi-hop questions ("which customers mentioned the bug that this PR fixed?") better than plain vector search, and answers can cite their sources.
Cognee is pluggable: it can use Neo4j, Kuzu or FalkorDB for the graph and pgvector, Qdrant, LanceDB and others for vectors, and it ships an MCP server so tools like Claude Code and Cursor can use it as memory. The Berlin company raised a €7.5M round in February 2026 led by Pebblebed and lists Bayer among its production users.
The Verdict
Who Should Use Cognee?
Best For
- Agents that need persistent memory across sessions
- Multi-hop questions over connected data
- Teams that want graph RAG without building the pipeline
- Regulated teams that need self-hosting
Not Ideal For
- Simple FAQ or single-document retrieval
- Teams without time for data modeling or ontologies
- Anyone wanting a finished end-user chat product
What's Great
- Graph plus vector retrieval beats vector-only RAG on connected questions
- Open-source core you can audit and self-host
- Official MCP server for coding agents
- Model- and storage-agnostic
- Very fast-growing community
Watch Out For
- Custom ontologies take data modeling effort
- Graph extraction adds LLM cost and processing time
- No polished end-user interface
- Overkill for simple retrieval
Pricing
View all features & details
Key Features
- Knowledge graph extraction from unstructured data
- Hybrid graph and vector retrieval
- Custom ontologies
- Connectors for Slack, Notion, Linear, Google Drive, GitHub
- MCP server
- Permissions and provenance
Integrations
- Neo4j, Kuzu, FalkorDB graphs
- pgvector, Qdrant, LanceDB vectors
- LangGraph, LangChain, LlamaIndex, CrewAI
- Claude Code, Cursor via MCP
How It Compares
| Feature | Cognee | Graphiti / Zep | Mem0 | Microsoft GraphRAG |
|---|---|---|---|---|
| Open source | Yes | Yes (Graphiti) | Yes | Yes |
| Knowledge graph | Yes | Temporal graph | Optional | Yes |
| Managed cloud | Yes | Zep Cloud | Yes | No |
| MCP server | Yes | Yes | Yes | No |
| Main focus | Agent memory and graph RAG | Changing facts over time | Personal memory | Global summaries |
Is this your tool? Add a badge