What is Cognee?
Cognee turns documents, code, and conversations into persistent memory that agents can retrieve across sessions. Text is organized into entities, relationships, and searchable chunks, while code is represented through symbols and dependencies. Session distillation can preserve accepted lessons as durable knowledge, and custom data models and ontologies can structure memory around an application's domain.
The platform runs locally as a Python package or through Docker, with a command-line interface, a browser-based interface, and an HTTP API. Agent connections include plugins for Claude Code, Codex, and OpenClaw, plus MCP access for clients such as Cursor and Cline. Cognee Cloud provides a managed deployment option.
Local text ingestion, retrieval, and session storage can run without an LLM key using downloaded extraction and embedding models. Generated answers and model-dependent media processing require additional configuration, while LLM-dependent improvement stages are skipped when no LLM is configured. The bundled GLiNER extractor is described as a demo, and the production-ready Postgres graph-store implementation is a licensed product.
Key capabilities
- Memory ingestion Converts text into entities, relationships, and searchable chunks, and code into symbols and dependency graphs.
- Context retrieval Selects graph, vector, or code context through automatic routing or an explicitly chosen search strategy.
- Session learning Distills accepted session lessons into permanent memory and applies feedback through memory improvement operations.
- Memory management Stores content in datasets or sessions and removes specific items or entire datasets.
- Agent access Exposes memory through Python and HTTP APIs, agent plugins, and an MCP server.
- Source synchronization Provides Google Drive OAuth and Gmail connectors, including shared-drive synchronization and progress indicators.
Where Cognee fits
Cognee fits into agent applications that need to retain project decisions, previous fixes, and learned rules between runs. It can also provide shared organizational memory connecting documentation, tickets, conversations, and code. Applications can inspect retrieved evidence before using it to answer questions or continue tasks.
Before deployment, teams should evaluate extraction and retrieval quality on their own data, confirm which workflows require configured models, and review provider calls when using hosted models. Self-hosted installations need authentication, dataset permissions, persistent storage, and compatible backend configuration. The single-user Docker demo disables access control and is not the configuration intended for an exposed production API.
Install
uv pip install "cognee[gliner]"Reviews of Cognee
Be the first to share your experience with Cognee.
Write a review
Loading review form…
