What Graphiti adds to MCP memory
Most memory servers store facts. Graphiti calls itself a framework for temporally aware knowledge graphs. Every piece of information you add with add_memory is an “episode”; entities become nodes and facts become edges, and fact searches can filter on valid_at and invalid_at. The graph updates incrementally and can answer historical questions. That is useful for agents that need to know what is true now, not just what was once said (“Anna leads the project now, Mark did before”).
The MCP server puts that graph behind a set of tools:
| Tool | Use |
|---|---|
| add_memory | Add an episode (text, message, JSON) |
| add_triplet | Add a fact as subject, relation, object |
| search_nodes / search_memory_facts | Find entities or facts |
| get_episodes / get_episode_entities | Read back what was added |
| delete_episode / delete_entity_edge / get_entity_edge | Inspect and remove facts |
| summarize_saga / build_communities | Summaries and clusters in the graph |
| clear_graph | Clear graph data for the selected group or groups |
| get_status | Health check |
What you need to run it
- Docker and Compose (the default starts FalkorDB in a container), or Python 3.10+ with Neo4j 5.26+.
- An LLM API key: OpenAI by default, or Anthropic, Gemini, Groq or Azure OpenAI. The embedding provider must be configured too; switching the LLM provider does not replace it automatically.
- A client: the server listens on
http://localhost:8000/mcp/by default and also supports stdio. Cursor and VS Code with GitHub Copilot connect over HTTP. For a local Claude Desktop setup, Graphiti supports stdio directly;mcp-remoteis an alternative for connecting to its local HTTP endpoint. Callget_statusafter starting to check the server and database are up.
Is it safe for a team?
Not as shipped, and the project says so itself. The Graphiti MCP page states that the server has no authentication, no authorization and no protection against prompt injection, and that you should not use it as a production deployment. Anyone who can reach the endpoint can read, change and delete the whole graph, and clear_graph is one of the tools. Graphiti does support group namespaces (group_id) and records which episode a fact came from, but the reference server does not enforce per-user authorization, so namespaces are no security boundary between colleagues.
Put that next to what the MCP specification’s security guidance asks for: tokens issued for that specific server, minimal scopes that grow step by step, and HTTPS. A team setup would need you to add all of that yourself: a reverse proxy with sign-in, per-person access, TLS, logging, and a decision about which tools to hide.
The misunderstanding
“It runs over HTTP, so the whole team can just point at it.” Technically yes. But the server does not check who is calling, so every person and every agent reaching it gets the same write and delete rights across groups, and a single prompt injection in any connected tool can reach everything it exposes. A shared endpoint is not the same as shared memory a team can trust.
When it does make sense
For one developer, or a research setup on a laptop, Graphiti MCP is a strong way to test time-aware memory in Cursor or Copilot. For comparison with other servers, see what an MCP memory server is, and for the full list of controls a company needs, is MCP safe to use in a company.
Where Bombos fits
Bombos is also a context graph with retrieval, but built to be shared: one memory for every AI tool your team uses (Claude, ChatGPT, Copilot, Grok, Cursor, Codex, your own agents), with team knowledge, learned rules and what already sits in mail, documents and repos, and a visible trail of what each answer rests on. It runs in our cloud or on your own server, and one person can try it first with their own agents.
Related questions
Does Graphiti MCP work with Claude Desktop?
Yes. For a local setup Claude Desktop can start the server over stdio; to reach its HTTP endpoint instead, the Graphiti docs point to a gateway such as mcp-remote. Cursor and VS Code with GitHub Copilot connect over HTTP directly.
Which database does Graphiti MCP need?
By default it runs FalkorDB in a Docker container. Neo4j 5.26 or later is the alternative.
Does Graphiti MCP need an OpenAI key?
It needs an API key for an LLM provider to extract entities and facts. OpenAI is the default; Anthropic, Gemini, Groq and Azure OpenAI are also supported.
Can I put Graphiti MCP on a public URL for remote colleagues?
The Graphiti docs warn not to expose the server to an untrusted network, because any caller can read, change or delete all graph data. You would have to add sign-in and access control in front of it yourself.
