How do I give Claude, ChatGPT and Cursor long-term memory with MCP?

Connect each app to the same memory server over MCP, and they all share one long-term memory. Add https://mcp.supermemory.ai/mcp as a remote MCP server in Claude, ChatGPT, Cursor or any other MCP client and sign in once in each. A decision you save from ChatGPT is then there when you ask Claude.

Below are the steps for each client, the tools the server gives each app, and when a plugin is the better fit. Everything is checked against our docs and Anthropic's and OpenAI's setup guides as of October 21, 2026.

Why don't Claude and ChatGPT share memory already?

Each app keeps its memory to itself. ChatGPT can draw on saved memories and past chats, and Claude has memory and chat search of its own, but neither can read the other's, and neither reaches Cursor. MCP, the Model Context Protocol, is an open standard for connecting AI applications to external systems. Connect every app to the same memory server and they share whatever gets saved there.

Two things don't happen on their own. Connecting an app doesn't copy its past chats into the server: only what gets saved through the tools, or uploaded, is shared. And two different MCP memory servers don't share anything just because they speak the same protocol, so every app has to connect to the same one.

How to connect each client

The server URL is the same in every client (setup):

https://mcp.supermemory.ai/mcp

Sign-in uses OAuth. The client opens supermemory in your browser, you sign in or create an account, and you choose which spaces the client can access. There's no API key to paste.

Claude (desktop and web)

1. Open Customize → Connectors, click + Add, then Add custom connector.
2. Name it Supermemory, paste the server URL, and click Continue. Keep the authentication settings Claude detects and click Continue again.
3. Under Authentication, choose Sign in now, keep the recommended OAuth client, and click Add.
4. Sign in to supermemory, choose the access you want (for example Read + Write or Full access), and click Authorize.

You can change or revoke access later from the same Connectors page (Anthropic's guide).

ChatGPT (web)

1. Open Plugins, select the plus button, then Add custom MCP server.
2. Name it Supermemory MCP and paste the server URL under Connection. Leave the authentication settings as they are.
3. Select I understand and want to continue, then Create as a plugin, and sign in with Supermemory when ChatGPT asks.
4. On the authorization page, pick Read + Write (or Read only for search alone), Full access or Scoped to chosen spaces, and how long the connection should last. Then select Approve (guide).

To make supermemory ChatGPT's default memory, the guide suggests turning off ChatGPT's own Enable memory setting, which doesn't delete your past chats, and adding this to your custom instructions:

Use Supermemory MCP as my default memory and workspace memory whenever its tools are available. Use it first for both recall and saving new memories.

Cursor

Add the server to ~/.cursor/mcp.json and complete the sign-in Cursor opens:

{
  "mcpServers": {
    "supermemory": {
      "url": "https://mcp.supermemory.ai/mcp"
    }
  }
}

For coding work, the Cursor plugin does more than MCP: it searches memory with every substantive prompt and saves each turn without waiting for a tool call. The Cursor memory guide covers both options.

Any other MCP client

Add the URL as a remote HTTP MCP server and complete the OAuth flow the client opens. Menu names and config file locations vary between clients, and clients that take a JSON config generally accept the mcpServers shape above (setup).

Once an app is connected, check it by asking "What Supermemory spaces can I access?" or "What is my active Supermemory space?"

What each app can do once it's connected

You ask in plain language and the assistant picks the tool. These are the tools the server exposes today (MCP docs):

Tool	What it does
search_memory	Searches one space for memories that match a question
get_profile	Returns stable and recent profile facts for a space
add_memory	Saves something to memory
forget_memory	Forgets something that's outdated
list_documents	Lists stored documents with their summaries
get_document	Reads one stored document
list_memories	Lists recent memories and the documents they came from
list_spaces	Lists the spaces you can access
who_am_i	Shows the signed-in account, its access and the active space

Clients that support MCP Apps also get four interactive views inside the chat: a space picker (select-space), a save form you can edit before anything is stored (guided-save), a file upload (upload-file), and an interactive memory graph (memory-graph). Some clients also expose a context prompt that attaches the profile for your active space and up to three recently active ones. More on the server itself is on the MCP page.

Spaces decide what each app sees

A space keeps a project's documents, memories and profile together, so recall doesn't mix unrelated work. Tools use your active space unless you name another one. Say "search the launch space for the pricing decision" for a one-off, or ask to switch spaces for everything after that (how spaces work).

Check the account and space in each app before you rely on shared memory. Being signed in with the same email in two AI products doesn't by itself mean both are connected to the same supermemory account or space, so ask each one "Which Supermemory account and space are you using?" For a team, decide which spaces are shared and which stay personal, and give each app access only to the spaces it needs.

Test it with one fact

Use a made-up fact that no model could infer:

1. In ChatGPT, say: "Save to Supermemory: the demo project's release codename is Cedar-47. This is test data."
2. Wait a little, since memories form in the background. Then ask Claude: "What's the demo project's release codename?" Check that the answer came from a supermemory search, not a guess.
3. Change it from either app ("the codename is now Cedar-48") and ask again.
4. Ask an app to forget the test fact, then search once more.
5. Ask the same question with a different space active. The codename shouldn't come up there.

This separates the things that can fail: the connection, the save, the space, and the answer.

MCP or a plugin?

Use MCP for chat apps, and a plugin for a coding agent that has one. The difference is who decides when memory gets used:

	MCP server	Coding agent plugin
Works in	Claude, ChatGPT, Cursor and any other MCP client	Claude Code, Codex, Cursor, OpenCode and other agents with a supermemory plugin
Recall	When the model decides to call a tool	Automatically, with each substantive prompt
Saving	When the model calls add_memory, often because you asked	Automatically, as you work
Where memories go	Your active space, or the space you name	One container per repository, shared by the plugins
Sign-in	OAuth, once in each client	Browser sign-in or an API key

With MCP, say it plainly when something matters ("save this to Supermemory"), because the model may not call a tool on its own. The plugins hook into the agent instead, so nothing waits on the model's judgment. The coding agents page lists the plugins, and the quickstart covers agents you build on the API. If you're building on the Vercel AI SDK, the AI SDK guide covers that path.

Troubleshoot at the failing step

Symptom	First check
The server won't connect	The URL, whether the client supports remote MCP servers, and the OAuth error it shows
Connected, but no tools show up	Whether the tools are enabled in the client and what access you approved
A save worked, but another app finds nothing	The account, the space, whether the memory has formed yet, and what the search actually returned
An old answer keeps coming back	Whether it came from memory or from the chat's own context, and whether the old fact was forgotten
Unrelated project context shows up	The active space and which spaces the app can access

Keep the actual error and use the client's supported sign-in flow. Turning off certificate checks or ignoring OAuth errors won't fix memory.

Common questions

Does connecting an app copy my old ChatGPT or Claude memories?

No. Only what gets saved through the tools, or uploaded, ends up in supermemory. For ChatGPT, the guide suggests opening Manage in its memory settings, copying the summary, pasting it into a chat with supermemory enabled and saying "Add this to Supermemory" (guide).

Do I need an API key?

No. The server uses OAuth, so each client signs in through your browser (setup).

Can I limit what an app can see?

Yes. Each client needs your approval before it can read or write memory (changelog). On the approval page, choose Scoped instead of Full access to pick specific spaces, or Read only if the app should only search.

Which clients does it work with?

Any client that can connect to a remote MCP server. Our docs have step-by-step guides for Claude and ChatGPT, and the setup page covers Cursor and other clients.