Working Memory exposes your notes, reminders, and search to AI assistants through the Model Context Protocol (MCP) — an open standard for connecting AI assistants to external tools and data. Once connected, an assistant can save thoughts to your account or look up things you've stored earlier without leaving the chat you're already in.
This article covers manual setup with a Personal Access Token for
clients that support custom MCP server URLs and header-based auth —
Claude Code, Cursor, Continue, Zed, and any other MCP-capable tool that
accepts an Authorization: Bearer header.
Claude Desktop is coming separately. Claude Desktop currently requires the OAuth flow, which is in active development. Stay tuned for updates there.
What the assistant can do
Two tools become available once you connect:
save_memory— saves a note, idea, fact, todo, or reminder. The AI parses the text and stores it as one or more structured items, scheduling any reminders automatically.search_memory— semantic search across everything you've saved. Useful when the user references something they stored earlier.
The assistant decides when to call these tools based on the conversation. You don't have to invoke them explicitly.
Set up
1. Create a Personal Access Token
- Sign in at app.workingmemory.ai.
- Open Integrations from the chat navigation.
- Click the connect button to mint a new token. In the dialog that appears, expand the manual setup section and click Reveal to see the token, then copy it. The token is shown only once — save it somewhere safe.
Token prefix: tokens start with wm_pat_ so they're easy to spot in config
files. Anyone with the token can read and write your Working Memory, so
treat them like passwords.
2. Configure your client
You need two values:
- Server URL:
https://app.workingmemory.ai/mcp - Authorization header:
Authorization: Bearer <your token>
The exact config format depends on which assistant you use.
Claude Code
Add an entry to ~/.claude/mcp.json (or your project's .claude/mcp.json):
{
"mcpServers": {
"workingmemory": {
"url": "https://app.workingmemory.ai/mcp",
"transport": "http",
"headers": {
"Authorization": "Bearer wm_pat_your_token_here"
}
}
}
}
Restart Claude Code; the save_memory and search_memory tools should
appear in the tool list.
Cursor
Settings → MCP → Add Server, with:
- Name:
workingmemory - Type:
SSE - URL:
https://app.workingmemory.ai/mcp - Headers:
{"Authorization": "Bearer wm_pat_your_token_here"}
Cursor's default type is Command (for local stdio servers); make sure you switch to SSE for this remote HTTP connection. The headers field expects a JSON object, not a raw header string.
Other clients
Any MCP-capable client supporting HTTP/JSON-RPC transport can connect.
Point it at the server URL and pass the token in an Authorization: Bearer
header.
3. Test the connection
In your AI assistant, try something like:
"Save a note: I need to renew my passport before October."
The assistant should call save_memory and confirm. You'll see the note
appear in your Working Memory account immediately.
Manage and revoke tokens
The Integrations page lists every token you've created with its last-used timestamp. To revoke a token (lost device, suspected leak, no longer using that assistant), click Revoke on the row. Any client using that token stops working immediately. Mint a fresh one if you still need access.
If you lose a token, revoke and re-mint — the original is shown only once and cannot be recovered.
Privacy and security notes
- Tokens scope to your account only. They never see other users' data.
- Working Memory does not train on or share the contents of
save_memorycalls. Your notes stay yours. - Tokens don't expire on their own, but you should revoke any you stop using. The Integrations page shows last-used timestamps to help you spot stale tokens.
Troubleshooting
- The assistant doesn't see the tools after I configured the server. Restart the assistant fully. Most clients only load MCP config at startup.
- Saves succeed but
search_memoryreturns nothing for a recent note. Search indexing runs a moment behind saves — wait ~30 seconds and try again. - Authorization errors. The token may have been revoked or mistyped.
Re-check the
Authorization: Bearer <token>header; the value must start withwm_pat_and have no extra whitespace.
Still stuck? Email support with the client name and the error message you're seeing.