The moving parts
The Model Context Protocol (MCP) is an open standard for connecting AI applications to outside systems: data sources, tools and workflows. Its makers compare it to a USB-C port for AI. Claude, ChatGPT, Visual Studio Code and Cursor all speak it.
A memory server is just one kind of MCP server. It does three things:
- Stores facts somewhere outside the model: a file, a graph database or a hosted account.
- Exposes tools the AI can call, like “add a memory” or “search memories”.
- Returns results that the client puts into the model’s context for that turn.
So the AI does not “remember” in the human sense. It decides to call a search tool, gets text back, and reasons with it. If it does not call the tool, the memory is not there.
Three common servers side by side
| Aspect | Official server-memory | Graphiti MCP (Zep) | Mem0 MCP |
|---|---|---|---|
| Storage | Local JSONL file | Graph database (FalkorDB or Neo4j) | Your Mem0 account, hosted |
| Runs where | On your machine, started by the client | Docker or Python, HTTP on localhost:8000 by default | https://mcp.mem0.ai/mcp |
| Example tools | create_entities, add_observations, search_nodes, read_graph | add_memory, search_memory_facts, get_episodes | add_memory, search_memories, delete_memory |
| Status per its own docs | Reference implementation, not production-ready | Experimental, no authentication | Needs a Mem0 Platform account and API key |
The official server stores a knowledge graph of entities, relations and observations, with nine tools. Graphiti adds time awareness: facts carry when they were true. Mem0 keeps memories in its cloud and signs you in through the browser or an API key.
How a memory call actually flows
- You tell Claude: “Remember that our invoices go out on the 25th.”
- Claude decides to call a tool: on the official server it creates an entity with
create_entities, or adds the fact to an existing entity withadd_observations. - The server writes that line to its file or database.
- Next week, in a new chat, you ask about invoicing. Claude calls
search_nodeswith “invoice”. - The server returns the stored fact, and Claude answers with it.
Step 4 is not automatic: the client must retrieve the stored fact and include it in the model’s context. The model has to think of searching. Many people add an instruction to their system prompt or project instructions that tells the AI to check memory first.
The common misunderstanding
People expect an MCP memory server to give “the AI” a memory. It gives one MCP client, with that server configured, access to one store. The official server, for example, writes to a file on your disk. Your colleague’s Claude does not see it. Your own ChatGPT does not see it either, unless you connect it to a server it can reach. The official README notes that each client starts its own server process; it uses a lock file so writes from several clients on one machine take turns, and warns that sharing is unsafe on a network file system that does not honor exclusive file creation.
Graphiti can be reached over HTTP, but its own page warns that it has no authentication or authorization: anyone who can reach it can read, change and delete everything.
What this means for a team
A memory server is personal unless you make it shared on purpose. To share it you need a remote server, sign-in per person, rules about who may write, and a way to see why the AI used a certain fact. Before sharing a store, check each server’s identity model, access controls and provenance features; MCP itself does not provide these. How memories are stored, selected and searched also depends entirely on the server. If you want the deeper setup steps, see how to set up server-memory in Claude, and for the risks read is MCP safe to use in a company.
Where Bombos fits
Bombos is a plugin that gives every AI tool a team uses (Claude, Claude Code, ChatGPT, Copilot, Grok, Codex, Cursor, your own agents) one shared memory. It holds team knowledge, rules the AI learned about how you work, and what already exists in mail, documents and repos, and it shows what each answer rests on. It runs in our cloud or on your own server, and one person can try it first for their own agents.
Related questions
Does an MCP memory server make the model itself remember?
No. The model stays the same. The server stores the data, and the model only sees a memory when it calls a tool such as search_nodes or search_memories and the result lands in its context.
Which MCP memory server is the official one?
The MCP project publishes a reference server called Knowledge Graph Memory Server (package @modelcontextprotocol/server-memory). The project describes its reference servers as educational examples, not production-ready solutions.
Can Claude, ChatGPT and Copilot share one MCP memory server?
Only if the server is reachable by all of them, which in practice means a remote server over HTTPS with proper authentication. A local stdio server on one laptop is only visible to the clients on that laptop.
Is an MCP memory server the same as Claude memory or ChatGPT memory?
No. The built-in memory of Claude or ChatGPT lives inside that product. An MCP memory server is a separate store that any MCP client can call, if you connect it.
