Connect Local models and custom agents.

The model can run on your laptop or in the cloud. Its agent application connects to the same hosted memory using MCP.

Browser sign-in · SDKs & custom hosts

Start in MarkdownMem

Sign in to the memory desk, select or create a memory store, and open Connect an agent. Copy the exact endpoint. The examples use placeholder workspace and store names.

Set up your client

Most clients that support remote MCP and browser sign-in need only your MCP URL and any name for the connection. Paste the URL, name it, and connect; your browser opens for sign-in. The required transport is Streamable HTTP.
For Python, install FastMCP, set MARKDOWNMEM_MCP_URL to your copied endpoint, and run the example below after creating /context.md in the memory desk. FastMCP opens your browser for authorization.
Sign in, review the requested access, and return to your client. Run the read check below before connecting a model to the returned tools.

import asyncio
import os
from fastmcp import Client

async def main():
    async with Client(
        os.environ["MARKDOWNMEM_MCP_URL"],
        auth="oauth",
    ) as client:
        result = await client.call_tool("memory_read", {"path": "/context.md"})
        # Pass result to your agent; avoid logging private memory.

asyncio.run(main())

A model alone does not implement MCP: its host or SDK does. Stdio-only clients need an HTTP bridge. A local model still uses the hosted service; for fully local storage, self-host the open-source engine. A headless job can instead use a store-scoped automation key in private process configuration; see the optional automation section in the client index.

Check that it reads the right memory

Create /context.md in the browser with a harmless project fact. In your agent, ask:

Use MarkdownMem to read /context.md. Tell me the project fact it contains. Don't change any notes.

To check portability, connect a second client to the same store and ask it to read that note. To check writes, ask one client to save a small test note and confirm it appears in the web editor. Keep the client's usual tool approval controls enabled.

Tell the agent when to use memory

Before working, search MarkdownMem for relevant context. Read the matching notes. After finishing, update durable decisions and preferences. Don't save credentials or whole conversation transcripts.

Add this to your client's project instructions if you want it to apply in future sessions. Connecting an MCP server does not automatically import built-in memory or make every conversation save notes.

If something does not connect

Copy the endpoint again: a different store path selects different memory.
Complete the client’s sign-in action and check your account still has access to the store.
For browser OAuth, remove any old Authorization header or expired key that would take precedence.
For cloud or remote sessions, configure the connection in that environment; local settings may not carry over.
If only the interactive editor is missing, check MCP Apps support. Text tools can work even when a host does not render the UI.

Browser OAuth acts with your account's current permissions. For a narrower, unattended connection, see store-scoped automation keys.

Current client references

Instructions reviewed September 6, 2026 against the sources below. These are configuration guides, not a claim that every client version, plan, or managed workspace has been certified.

Connect an agent