# Cognee

> Open-source AI memory platform that turns data into a queryable knowledge graph, giving AI agents persistent long-term memory across sessions.

Cognee is an open-source AI memory platform built by Topoteretes UG, a Berlin-based company. It ingests data in any format, builds a self-hosted knowledge graph, and lets AI agents recall context persistently across sessions. The core SDK is licensed under Apache 2.0 and available on GitHub, while a managed cloud offering (Cognee Cloud) handles infrastructure for teams that prefer not to self-host.

## What It Is

Cognee sits in the agent memory and GraphRAG category. Its core job is to replace ad-hoc vector retrieval with a richer memory layer that combines vector embeddings, graph relationships, and cognitive-science-grounded ontology generation. The result is a system where agents can not only find semantically similar content but also traverse connected relationships across documents, sessions, and data sources. The product exposes four primary operations — `remember`, `recall`, `forget`, and `improve` — through a Python SDK, an HTTP API, and an MCP server, so any MCP-compatible agent can read and write Cognee memory without custom integration code.

## Architecture and Storage Backends

Cognee is designed to run locally, on self-hosted infrastructure, or on Cognee Cloud. Under the hood it supports multiple graph and vector storage backends:

- **Graph stores**: Kuzu, NetworkX, Neo4j, FalkorDB
- **Vector stores**: LanceDB, Qdrant, Milvus, Redis
- **Relational metadata**: handled in-process

The product page also describes a Cognee Rust Engine — an embedded AI memory engine that bundles vector store, graph store, relational metadata, and local embeddings into a single binary with no external services required. The page claims a cold-start-to-query-ready time of approximately 350 ms, which the team says makes it viable for serverless functions, CLI tools, mobile apps, and edge devices.

## Integration Surface

Cognee ships first-party integrations for Claude Code, Cursor, Codex (OpenAI Agents SDK), LangGraph, CrewAI, and OpenClaw, plus a generic MCP server. The `@cognee.agent_memory` decorator wraps any async agent entrypoint and automatically composes graph memory and session memory, persisting the agent's execution history as queryable memory. Data source connectors for Slack, Notion, and Google Drive are available on higher-tier cloud plans.

## Open-Source Lineage and Current Status

The repository was created in August 2023 and is actively maintained. The GitHub page reports 18,306 stars and 1,949 forks as of mid-2026. The latest release is **v1.2.0**, published on June 21, 2026, and the homepage notes that "Cognee v1 is coming soon," signaling a major version milestone in progress. The project is part of the UC Berkeley RDI Xcelerator program, according to the homepage. The vendor's homepage states the SDK processes over 5 million runs per month.

## Deployment Paths

Teams can reach production through several routes:

- **Local / self-hosted**: `pip install cognee`, configure an LLM API key, run the pipeline directly or via the CLI (`cognee-cli`)
- **Cognee Cloud**: managed serverless hosting; connect any Python agent with `await cognee.serve(url=..., api_key=...)`
- **One-click PaaS**: Modal, Railway, Fly.io, Render, and Daytona configurations are included in the repository's `distributed/` folder
- **On-premises / air-gapped**: the product page states full GDPR compliance, data encrypted at rest and in transit, and support for air-gapped enterprise deployment

## Why It Matters for Agent Builders

Standard RAG retrieves similar text chunks; Cognee adds graph traversal so agents can answer multi-hop questions that require connecting facts across sources. The vendor publishes benchmark comparisons against LightRAG, Graphiti, and Mem0 on human-like correctness, DeepEval F1, and DeepEval EM metrics. A research paper — "Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning" (Markovic et al., arXiv 2505.24478, 2025) — accompanies the project. The open-source core means teams can start locally at no cost and migrate to managed cloud or enterprise deployment as workloads grow.

## Features
- Persistent long-term agent memory across sessions
- Knowledge graph construction from any data format
- Vector + graph hybrid search (GraphRAG)
- Four-verb memory API: remember, recall, forget, improve
- MCP server for any MCP-compatible agent
- Python SDK, HTTP API, and CLI
- Session memory with automatic graph sync
- @cognee.agent_memory decorator for async agents
- Custom ontology generation
- Multi-tenant isolation
- Permissions and access control
- Data source connectors (Slack, Notion, Google Drive)
- Support for multiple graph backends (Neo4j, Kuzu, NetworkX, FalkorDB)
- Support for multiple vector backends (LanceDB, Qdrant, Milvus, Redis)
- Cognee Rust Engine: embedded single-binary memory engine
- Air-gapped and on-premises deployment
- GDPR compliance with data encryption at rest and in transit
- OTEL collector and audit trails
- Self-improving memory from feedback
- Autoscaling compute on Cognee Cloud

## Integrations
Claude Code, Cursor, Codex (OpenAI Agents SDK), LangGraph, CrewAI, OpenClaw, Continue, Cline, Zed, Gemini CLI, Windsurf, Aider, Goose, Roo Code, Slack, Notion, Google Drive, Neo4j, Kuzu, NetworkX, FalkorDB, LanceDB, Qdrant, Milvus, Redis, Modal, Railway, Fly.io, Render, Daytona

## Platforms
API, CLI, DEVELOPER_SDK, WEB

## Pricing
Open Source, Free tier available

## Version
v1.2.0

## Links
- Website: https://www.cognee.ai
- Documentation: https://docs.cognee.ai/
- Repository: https://github.com/topoteretes/cognee
- EveryDev.ai: https://www.everydev.ai/tools/cognee