LLMemory MCP Server
A remote MCP server that lets AI coding agents search and pull context from your imported AI chat history (ChatGPT, Claude, Codex, Claude Code) without copy-pasting. Reads your vault live from your own Google Drive. Zero server-side storage, fully private.
Server URL: https://mcp.llmemory.xyz/mcp
Transport: Streamable HTTP
Auth: OAuth 2.1 with Dynamic Client Registration (no API keys β browser flow on first use)
Install in Cline
This is a remote MCP server β no cloning, no npm install, no local process. Cline just needs the URL.
Option A β Cline Dashboard (recommended):
- Log in to the Cline Dashboard β Organization β MCP Servers
- Add a new Remote MCP server
- Name:
LLMemory Β· URL: https://mcp.llmemory.xyz/mcp
- Save. In VS Code/JetBrains, open the Cline panel β MCP Servers β connect
llmemory.
- On first tool call, a browser window opens β authorize with Google and grant Drive access.
Option B β Cline MCP config JSON:
Add this to your Cline MCP settings (cline_mcp_settings.json):
{
"mcpServers": {
"llmemory": {
"type": "remote",
"url": "https://mcp.llmemory.xyz/mcp"
}
}
}
Restart the Cline extension. On first tool call, the OAuth browser flow opens automatically.
Install in other MCP clients
Same URL works in any MCP client that supports Streamable HTTP + OAuth 2.1:
- Cursor β LLMemory Cursor plugin
- VS Code / GitHub Copilot β add to
.vscode/mcp.json: { "servers": { "llmemory": { "type": "http", "url": "https://mcp.llmemory.xyz/mcp" } } }
- Claude Desktop / Claude Code β
claude mcp add --transport http llmemory https://mcp.llmemory.xyz/mcp
- Open WebUI β Admin Settings β External Tools β add Streamable HTTP MCP server with URL
https://mcp.llmemory.xyz/mcp
| Tool | Purpose |
|---|
list_sources | Show which AI providers have data in your vault and counts |
list_recent | Recently updated conversations (id, title, source, message count). Filter by source |
search_chats | Full-text search across all imported conversations. Returns snippets + conversation IDs |
get_conversation | Fetch one full transcript by ID (Markdown or JSON) |
get_context_pack | Agent-ready briefing (goal, recent state, files, open tasks) for one conversation or a whole folder |
What it reads
LLMemory reads the LLMemory/ folder in your Google Drive β a manifest.json plus one JSON file per conversation, written by the LLMemory browser extension or CLI. The server is read-only and zero-storage: it brokers Google OAuth, streams the result, then forgets everything. Your chats never leave your Drive on a server.
Privacy
- Zero server-side storage. No chat content, no per-user database. Cloudflare Workers free tier.
- Your Drive is the storage. Chats live in your own Google Drive
LLMemory/ folder.
- Read-only. The connector never writes to your vault.
- OAuth-scoped. The server only ever sees your
LLMemory/ Drive folder (drive.file scope).
Full privacy posture: STORE_PRIVACY.md
Examples
- "What did I figure out about the OAuth worker last week?" β searches your vault
- "Pull my most recent Codex session" β
list_recent filtered to codex, then get_conversation
- "Give me a context pack for the session where we built the Claude connector" β
get_context_pack
- "Show me which chats have data imported" β
list_sources
Links
License
MIT