Cognee

Overview

Cognee is an open-source agent memory platform for LLM agents. It builds persistent memory across sessions from documents, code, and conversations, turning them into a self-hosted knowledge graph that agents can search and reuse. It supports graph, vector, and relational retrieval and can run locally, in Docker, on-prem, or on Cognee Cloud.

Key Features

• Persistent long-term memory across sessions for AI agents
• Graph, vector, and relational retrieval
• Self-hosted knowledge graph engine
• First-party integrations for Claude Code, Cursor, LangGraph, OpenClaw, and more, plus an MCP server
• Custom ontologies, custom data models, and permissions control
• Local use without an OpenAI or Anthropic API key

Use Cases

• Build a Company Brain by bringing documentation, conversations, tickets, code, and agent work into shared memory.
• Give agents memory across runs so they retain project context, past decisions, fixes, and learned rules.
• Ground agents in your domain by structuring memory around entities and relationships with custom data models and ontologies.
• Support customer-facing agents that retrieve cited facts, follow domain rules, and improve from real usage.

Getting Started

• Requires Python 3.10–3.14.
• Install with uv pip install "cognee[gliner]" or another Python package manager.
• Optionally configure the LLM by setting LLM_API_KEY or creating a .env file from the template.
• Run locally without an LLM by saving the quickstart script and running python quickstart.py, or use cognee-cli remember and cognee-cli recall.
• Explore the bundled demo with cognee-cli demo.

Deployment & Requirements

• Requires Python 3.10–3.14.
• Can run self-hosted, in Docker, on-prem, or on Cognee Cloud.
• Cognee can be deployed BYOC in your own cloud.
• Cognee Cloud is a managed service option.

Before You Adopt

• License: Apache-2.0. Review its terms before using, modifying, or distributing the project.
• Default processing uses OpenAI for language models and embeddings, and processing plus generated answers make provider calls.
• Generated answers and media processing that requires vision or transcription models need additional LLM configuration.