What it's for and how it works
A normal AI forgets everything at the end of the conversation: its "memory" lives only inside the open chat. The memory server breaks this limit. It keeps a knowledge graph (a network of facts linked to each other: people, projects, preferences and the connections between them) saved as a file on your computer. When you say "remember that I work in the wine industry", the AI writes it there; in the next chat it re-reads it and takes it into account.
The memory is yours and local: it's a file on your disk, not something that lives inside the AI's app. You can read it, save it, move it. That's the difference that makes it reliable: you know where it is and what it contains.
How to do it
From any assistant that supports MCP, the principle is the same.
- Add the server to the configuration file. The working syntax:
{
"mcpServers": {
"memory": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-memory"]
}
}
}
This is the official reference memory server: it creates and manages the knowledge graph on its own.
- Restart the assistant completely, so it loads the server.
- Teach it the facts to remember. Write to the AI explicitly: "remember that my main client is a winery in Tuscany and that I prefer a formal tone in emails". The AI saves them in the graph.
- Check the recall. In a new chat ask: "what do you remember about me and my clients?". If it lists the facts, the memory works.
A concrete example
Giorgio uses AI to write the communications for his travel agency. Every time he repeated the same premises: what kind of clients he has, the tone he wants, the destinations he handles. He connects the memory server and, in the first conversations, tells the AI to remember these points.
From then on he opens a new chat and just writes "draft me the email for a client interested in Japan". The AI already knows what tone to use and what kind of trips he offers, because it re-read the memory. Giorgio has stopped pasting the usual block of context. The graph grows over time: the more he uses it, the more the assistant "knows" him.
When it does NOT work (and how to fix it)
If the AI doesn't recall what it should remember
Sometimes the assistant doesn't consult the memory on its own initiative. Tell it openly at the start of the chat: "check what you know about me in the memory before answering". Over time you'll learn which assistants use it automatically and which need prompting.
If wrong or old memories pile up
The memory can fill up with outdated information (a client you no longer follow). Ask the AI to "update the memory" or to "forget" a specific fact. Every now and then it's worth asking it to list what it has saved, so you can correct what's no longer valid.
If you don't know where the data ends up
The saved facts sit in a file on your computer, not on the AI's cloud. If you want to see or save that file, it's in the folder where the server creates it: ask the assistant itself where it keeps it. Knowing where it is also lets you take it elsewhere.
A tip from someone who actually uses it
Keep the memory essential and curated, not a warehouse of everything. Save the stable facts that recur (who you are, how you work, your regular clients), not the details of a single job. A clean memory makes the AI more useful; one full of old stuff confuses it, like a desk buried under papers.
Frequently asked questions
What's the difference from the memory already built into ChatGPT?
The built-in memory of some apps lives inside that platform and you don't really control it. The one via MCP is a file of yours, on your computer: you read it, edit it, move it between different assistants. It's the difference between a rented memory and one you own.
Can I use the same memory with different AIs?
In theory yes, because it's a file and MCP is a standard: the same memory server can be connected to different assistants. It's one of the advantages over memories locked inside a single app, which stay there and don't move.
Does external memory replace a context document of mine?
Here's the misunderstanding to clear up: memory via MCP is convenient, but the most solid portability remains a context document written by you (a sheet with who you are, how you work, your projects). The memory server is a useful automation; the document is the source you control one hundred percent and that works everywhere, even with an AI that doesn't speak MCP.