MCP
Package: @turbomem/mcp
A Model Context Protocol server built on turbomem. It gives Claude a private, local long-term memory: Claude can remember durable facts about the user and recall them in future conversations. Everything is stored on disk on the user's machine (the same ~/.turbomem/data store as the CLI); nothing is uploaded anywhere except the AI provider chosen to extract and embed the notes.
It ships as a one-click Claude Desktop extension (.mcpb), so non-technical users never touch a terminal or config file.
Install in Claude Desktop (no terminal)
- Download
turbomem.mcpb(from the project Releases). - Double-click the file - Claude Desktop opens an install screen. (You can also drag it onto the window, or use Settings → Extensions → Advanced settings → Install Extension…)
- Fill in two fields and click Install:
- AI provider - type
openai,google, oranthropic(leave asopenaiif unsure). - API key - paste the key for that provider.
- AI provider - type
Then just chat. Claude saves and recalls memories on its own. Try "Remember that my dog is named Rex," then, in a new chat, ask "What's my dog's name?"
Which API key do I need?
You only need one key, matching the provider you chose:
| Provider | Get a key | What it powers |
|---|---|---|
openai | platform.openai.com | Understanding + searching memories |
google | aistudio.google.com | Understanding + searching memories |
anthropic | console.anthropic.com | Understanding memories (extraction) |
openai and google use a single key for both understanding and search. See the Providers reference for details.
Using Anthropic (Claude)?
Anthropic can organize your memories, but it has no embedding model, so memory search needs a second provider. On the install screen, set:
- AI provider →
anthropic, API key → your Anthropic key - Search provider →
openaiorgoogle, Search API key → that provider's key
That runs extraction on Claude and search on OpenAI/Google over HTTP — no local model or extra download required. (Advanced: you can instead run search fully on-device by installing @huggingface/transformers via the manual setup.)
Tools
Claude is instructed to use these automatically, but you can also ask for them directly ("list everything you remember about me", "forget that I like dark mode").
| Tool | Description |
|---|---|
remember | Saves durable facts about the user to local memory |
recall | Searches memory for anything relevant to a query |
list_memories | Lists everything currently stored |
forget | Deletes one memory by id |
forget_everything | Erases all memories (only on explicit confirmation) |
Advanced: manual setup
For another MCP client, or to wire it up by hand, run the server over stdio.
npm install -g @turbomem/mcpAdd it to claude_desktop_config.json:
{
"mcpServers": {
"turbomem": {
"command": "turbomem-mcp",
"env": {
"TURBOMEM_PROVIDER": "openai",
"TURBOMEM_API_KEY": "sk-...",
"TURBOMEM_USER_ID": "me",
"TURBOMEM_DATA_DIR": "/Users/you/.turbomem/data"
}
}
}
}To run extraction on Anthropic and search on OpenAI (no local model):
{
"mcpServers": {
"turbomem": {
"command": "turbomem-mcp",
"env": {
"TURBOMEM_PROVIDER": "anthropic",
"TURBOMEM_API_KEY": "sk-ant-...",
"TURBOMEM_EMBEDDINGS_PROVIDER": "openai",
"TURBOMEM_EMBEDDINGS_API_KEY": "sk-..."
}
}
}
}Or run without installing:
npx @turbomem/mcpTo run Anthropic search fully on-device instead (no embeddings key), set TURBOMEM_EMBEDDINGS_PROVIDER=local and install the local model:
npm install -g @huggingface/transformersEnvironment variables
| Variable | Default | Notes |
|---|---|---|
TURBOMEM_PROVIDER | openai | Extraction provider: openai, google, or anthropic |
TURBOMEM_API_KEY | - | API key for the extraction provider (required) |
TURBOMEM_EMBEDDINGS_PROVIDER | mirrors provider | Search provider: openai, google, or local |
TURBOMEM_EMBEDDINGS_API_KEY | - | API key for the search provider (needed unless it is local) |
TURBOMEM_USER_ID | me | Label for this memory profile |
TURBOMEM_DATA_DIR | ~/.turbomem/data | Where memories are stored on disk |
TURBOMEM_EXTRACTION_MODEL | provider default | Override the fact-extraction model |
TURBOMEM_EMBEDDING_MODEL | provider default | Override the embedding model |
Standard provider env vars (OPENAI_API_KEY, GEMINI_API_KEY, ANTHROPIC_API_KEY) are also honored as fallbacks for the matching provider.
Build the .mcpb yourself
pnpm --filter @turbomem/mcp pack:mcpbThis builds the server, installs its production dependencies into a bundle, and writes turbomem.mcpb in the package directory. Double-click it to install, or submit it to the Claude Connectors Directory so users can install it from inside Claude Desktop.