Cognee

What topoteretes/cognee MCP server does

The topoteretes/cognee MCP server provides an agent-facing path into Cognee, an open-source memory platform. It is intended for applications that need more than a single conversation’s context: documentation, conversations, tickets, code, decisions, and session lessons can be brought into a shared memory system.

Cognee supports persistent memory and session memory. Persistent content is processed into a connected knowledge representation, while session-scoped content can be recalled quickly and synchronized with longer-lived graph memory in the background. This makes the topoteretes/cognee MCP server relevant to agents that need to retain project context between runs or connect facts across different source types.

The supplied material describes MCP as one way to connect an existing agent, but does not list the MCP server’s individual tool names, transport details, or request schemas.

How it works

Cognee accepts text, code, and session guidance through its memory workflow. Text is converted into entities, relationships, and searchable chunks. Code is modeled around symbols and dependencies. Session distillation selects useful lessons for inclusion in permanent memory rather than treating every temporary interaction as durable knowledge.

At query time, retrieval can select graph, vector, or code-oriented context. Automatic routing chooses a search strategy, while the underlying application can inspect the returned evidence before using it to answer a question or continue an agent task. The core operations described by the project are remembering content, recalling context, improving or enriching memory, and forgetting an item or dataset.

Setup and configuration

Cognee supports Python 3.10 through 3.14 and can be installed with pip, uv, or another Python package manager. The repository also provides a demo that can run without an API key.

For the default language-model and embedding configuration, set LLM_API_KEY to an OpenAI API key. Processing and generated answers make provider calls. The project points to alternative provider configurations and local Ollama models, but the provided material does not specify their complete environment-variable requirements.

The README identifies plugins and MCP as agent connection options. It specifically documents integrations for Claude Code, Codex, OpenClaw, Cursor, and Cline, although the excerpt does not provide a direct launch command for the MCP server itself. Configure the memory service according to the selected agent integration and deployment mode.

Tools and capabilities

The documented capabilities include:

• Store content permanently or associate it with a session.
• Recall relevant context using automatic routing or a selected search approach.
• Improve memory with enrichment, feedback, and session-to-graph bridging.
• Delete an individual memory item or dataset.
• Build knowledge from text, code, and other supported data sources.
• Use graph, vector, and code-aware retrieval paths.
• Define custom data models and ontologies for domain-specific entities and relationships.

The topoteretes/cognee MCP server should therefore be evaluated as a memory and retrieval layer, not as a general-purpose data connector. The provided description mentions ingestion from more than 30 data sources, but does not enumerate those sources in the available excerpt.

Limitations and notes

A provider key may be needed for normal language-model and embedding processing, even though the bundled demo can run without one. The default provider is OpenAI, and the project states that provider calls are involved in processing and generated answers.

The available material does not establish a hosted MCP endpoint, a dedicated package name for this server, a complete list of MCP tools, or a specific MCP transport. It also does not state a license name in the excerpt, so those details should be confirmed from the repository before deployment.

The topoteretes/cognee MCP server is a fit when persistent, connected memory is central to the agent workflow. For a simple one-off search over a small document set, its graph-building and session-memory model may be more infrastructure than necessary.