Cognee

What topoteretes/cognee does

topoteretes/cognee exposes Cognee’s memory engine through the Model Context Protocol. Agents can save text, ingest files, search stored information, remove selected data, or clear memory owned by the current setup. The memory model has two paths: session-aware entries can be kept in a fast cache, while entries created without a session identifier become persistent graph memory.

The server also includes code graph search for indexed code datasets. This supports questions and structural queries involving symbols, dependencies, architecture, change impact, and related code relationships. The memory and code-search functions make it suitable for knowledge and memory workflows where an agent must retain information across interactions or inspect an indexed codebase.

How it works

Use remember to provide text through data, or upload a file through filename and content_base64. Files can be up to 10 MB. Supplying session_id routes text-only data to the session cache; omitting it stores the data as permanent graph memory. Ingestion may run in the background when background=true is used.

recall searches memory and can use a session identifier to search the session cache first before falling back to the permanent graph. Reference information can be included in results, which makes it possible to identify individual data items for later deletion. forget can remove a dataset by name or ID, delete one item when given its dataset and data IDs, or remove all owned memory with everything=true.

The server supports stdio by default, plus SSE and Streamable HTTP transports. Streamable HTTP is presented as the recommended choice for web deployments. It can operate locally, connect to Cognee Cloud through a service URL, or use an existing Cognee FastAPI server.

Setup and configuration

Clone the Cognee repository, enter its cognee-mcp directory, install uv, and install the development dependencies with uv sync --dev --all-extras --reinstall. Activate the resulting virtual environment before starting the server.

A basic configuration uses an OpenAI key in .env:

LLM_API_KEY="YOUR_OPENAI_API_KEY"

Start the default stdio service with python src/server.py. SSE can be selected with --transport sse. For Streamable HTTP, use --transport http together with a host, port, and MCP path, such as 127.0.0.1, 8000, and /mcp.

Docker usage is also supported. The project provides a Dockerfile and publishes a cognee/cognee-mcp:main image. Container runs can select stdio, sse, or http through TRANSPORT_MODE. Optional dependency groups include database, storage, model-provider, document-processing, scraping, and tracing integrations.

Tools and capabilities

• remember stores text or base64-encoded files, with session and background-ingestion options.
• recall searches session and permanent memory with optional search controls and references.
• forget deletes individual items, datasets, or all owned memory.
• cognify_status reports progress for background ingestion jobs.
• code_search queries indexed code graphs through operations including query_facts, explore, traverse, find_path, impact_analysis, insights, architecture, and delta.
• Code navigation operations include capabilities, search, neighbors, subgraph, flows, hubs, and ask.

Limitations and notes

Session-cache storage accepts text rather than uploaded files. File ingestion through remember is limited to 10 MB per file. cognify_status is not advertised among the default tools, but it can be discovered through search_tools and called by name. A normal setup needs an LLM key, while hosts that provide MCP sampling can delegate completions to the host model; embeddings still require a provider.