An MCP memory server
MCP is a standard way for an assistant to reach a tool. Memory turns out to be close to the ideal thing to put behind it, for a reason worth spelling out.
What the protocol does
Before MCP, connecting a tool to an assistant meant a different integration per assistant, written by whoever owned the assistant. MCP inverts that: the tool describes what it can do, and any client that speaks the protocol can use it.
For a memory that matters more than for most tools. The value of a memory is proportional to how much of your working life passes through it, so a memory reachable from one assistant is worth a fraction of the same memory reachable from all of them.
Why memory suits it
Most MCP servers are read-only against a system that already exists: query a database, fetch a page, list files. Memory is different because the assistant writes to it, and what it writes is useful to a different assistant tomorrow.
That makes the store the shared thing and the assistants interchangeable. Claude records a decision on Monday, Cursor recalls it on Wednesday, and neither knows the other exists.
Connecting one
A hosted server is a URL and a transport in a config file. There is deliberately no API key in that file: authorisation happens through the browser with OAuth, so a copied config is not a copied credential and revoking is a button rather than a key rotation.
Claude Desktop, Claude Code, Cursor, Windsurf, Cline and ChatGPT all take the same block, give or take the wrapper key. The setup for each is on its own page.
What to check before trusting one
Whether it authorises per user or per installation. A server that takes one API key in a config file gives everyone who can read that file the same access, which is fine on your laptop and wrong on a shared machine.
Whether recall respects where a memory came from. A memory captured from a private channel should not become readable because an assistant summarised it, and this is the failure that is easiest to ship and hardest to notice.
Whether it tells you when retrieval is degraded. A server that silently answers from a narrower index looks identical to one that is working.
Common questions
- What is MCP?
- The Model Context Protocol, an open standard for connecting assistants to tools and data. A server describes what it offers and any compatible client can use it, instead of every assistant needing its own integration.
- Do I need an API key for an MCP memory server?
- Not for this one. The connection authorises through your browser using OAuth, so the config file contains a URL and nothing secret.
- Can several assistants share one memory server?
- Yes, and it is the main reason to use one. What you record in Claude is available in Cursor, because both are talking to the same store rather than each keeping their own.
- Does it work with local models?
- Any client that speaks MCP can connect, regardless of which model is behind it. The protocol is between the client and the server; the model is the client's business.