What is an MCP memory server?
An MCP memory server gives AI assistants tools to save and recall memories over the Model Context Protocol (MCP), the open standard that Claude, ChatGPT, Cursor, VS Code and most other assistants use to connect to outside tools. The memory lives in the server rather than in any one assistant, so what one saves, the others can find.
In short
- MCP is how an assistant connects to tools. A memory server is a tool for saving memories and searching them.
- A local server runs on your computer, for the apps on that computer. A remote server is a URL, so browser-based assistants can use it too, and every assistant shares one memory.
- Look for OAuth sign-in instead of keys in config files, a small set of tools, search by meaning, handling of duplicates and outdated facts, and a way for people to read and fix what is stored.
- GuardianMemory is a remote MCP memory server: one URL and two tools, free while we learn.
How an MCP memory server works
MCP splits the work in two. The assistant, which MCP calls the client, runs the conversation. The server offers tools, each with a name, a description and the arguments it takes, and the model decides when to call them.
A memory server's tools save and search. GuardianMemory offers exactly two:
| Tool | What the assistant does with it |
|---|---|
create_memory | Saves one lasting fact as a plain sentence, with optional tags. A near-duplicate is merged, and a fact that replaces an old one retires it. |
search_memory | Finds memories by meaning rather than keywords. With no query it returns your profile: pinned memories, then the most used, then the newest. |
When an assistant connects, the server also sends instructions: read the profile at the start of a conversation, search before asking the user to repeat themselves, save lasting facts one at a time, and never store secrets. So you tell Claude on Tuesday that the launch is on 3 November, ask Cursor on Thursday to put the launch date on the banner, and Cursor finds it.
Local or remote?
Memory servers come in two shapes, and the difference decides which of your assistants can share a memory.
| Local memory server | Remote memory server | |
|---|---|---|
| Where it runs | On your computer, started by each app | On the internet, at a URL |
| Which assistants can use it | Desktop apps and editors on that computer | Any MCP client, including web apps such as claude.ai and ChatGPT |
| One memory for every app | Only if every app is pointed at the same files | Yes: every app connects to the same URL |
| Other devices | No | Yes |
| Signing in | Nothing, or keys in config files | OAuth: approve each app once in the browser |
| Upkeep | Yours: installs, updates, backups | The provider's |
A local server keeps everything on your machine, which some people want. The cost is that the memory only exists where the server runs, and assistants that live in a browser cannot reach it.
What to look for in an MCP memory server
- A URL, not an install. Every assistant you use, including the ones in a browser, can connect to it.
- OAuth sign-in. You approve each app once, nothing secret sits in a config file, and you can disconnect one app without touching the rest. GuardianMemory lets clients register themselves and requires PKCE.
- A small set of tools. Models use tools more reliably when there are few and each one's purpose is obvious. GuardianMemory has two, and deliberately no delete.
- Search by meaning. Keyword matching will not connect “what database does billing use?” with “The billing service runs on Postgres and Stripe”. Embeddings do.
- An answer to the cold start. Search only helps an assistant that knows what to ask. A profile loaded at the start of each conversation gives it the basics first.
- Hygiene. Near-duplicates merged, outdated facts retired, newer facts favoured. Memory written by agents rots without it: why agent memory goes bad.
- A place for people. Read, edit, pin and delete every memory, see which assistant saved or searched what, revoke an app at once, and export the lot.
- Sharing with a wall. A memory your team shares, which nothing personal can reach by accident: see company brain.
Connect GuardianMemory to your assistants
Every assistant takes the same endpoint. The first time it connects, a browser opens so you can approve it.
https://app.guardianmemory.com/mcp
| Assistant | Where it goes |
|---|---|
| Claude | Settings → Connectors → Add custom connector, on claude.ai or in the desktop app |
| ChatGPT | Settings → Connectors → Create, then choose OAuth. Custom connectors need developer mode switched on |
| Claude Code | The command below, then /mcp to sign in |
| Cursor, Windsurf and other editors | An mcp.json with an mcpServers block, below |
| VS Code | .vscode/mcp.json, with a servers block |
claude mcp add --transport http guardian https://app.guardianmemory.com/mcp
{
"mcpServers": {
"guardian": { "url": "https://app.guardianmemory.com/mcp" }
}
}Every step for each assistant, including API keys for clients that cannot do OAuth, is in Connect an assistant. Then add a line to its instructions so it checks its memory every time.
Questions people ask
Do I need a separate memory server for each assistant?
No. A remote memory server is one URL, and every assistant connects to the same one. That is the point: what one assistant saves, the others can find.
Can ChatGPT use an MCP memory server?
Yes. ChatGPT connects to remote MCP servers as custom connectors, under Settings, Connectors, Create. Custom connectors need developer mode switched on.
Do I need an API key to connect?
Not for assistants that support OAuth, which is most of them: you approve each one once in your browser. API keys are for scripts and clients that cannot do OAuth.
What is the difference between a memory server and a knowledge graph?
A knowledge graph stores entities and the relations between them, which suits structured facts but asks the model to keep that structure tidy. GuardianMemory stores plain sentences and searches them by meaning, so a model only has to write one clear sentence per fact.
Who can see what is in my memory server?
Your personal bank is readable only by you and the assistants you approve. Memory text is sent to Google Gemini to make it searchable and to TypeSafe's Jev classifier to check whether a new fact replaces an old one, and nowhere else. What is stored, and where it goes.
One URL for every assistant
Connect Claude, ChatGPT, Cursor and the rest to the same memory in about a minute each.
Free while we learn. No card.
We are letting people in a few at a time.