Shared memory for LangChain & LangGraph agents

Why LangGraph teams add threadctx

• MCP is the interface LangChain already built for this. LangGraph's official langchain-mcp-adapters package connects any agent to any MCP server, and LangChain's Fleet product documents support for extending with "any MCP server." threadctx doesn't need a bespoke LangChain integration — it's already speaking the protocol LangChain built for exactly this.
• It bridges outside LangGraph too. threadctx is one server behind every client. A teammate's LangGraph agent writes a memory with memory_write, and a teammate running Claude Code or Cursor can recall it with memory_query — cross-framework sharing that nothing framework-specific can do on its own.
• Two tools, not a platform to learn. memory_write(content, tags?) and memory_query(task_description, max_results?) bind into your graph like any other LangChain tool — call memory_query before a node that does risky or repeated work, and memory_write after one that resolves something worth remembering.

Add it to a LangGraph agent

langchain-mcp-adapters is LangChain's own package for connecting agents to MCP servers over stdio or HTTP. For local mode, drop the env block — no key needed.

from langchain_mcp_adapters.client import MultiServerMCPClient

client = MultiServerMCPClient({
    "threadctx": {
        "command": "npx",
        "args": ["-y", "threadctx-mcp"],
        "env": {
            "THREADCTX_MODE": "cloud",
            "THREADCTX_API_KEY": "tctx_...",
        },
        "transport": "stdio",
    }
})

tools = await client.get_tools()  # memory_write, memory_query — ready to bind to your graph

Same idea if you're driving threadctx from LangChain's Fleet instead of raw LangGraph — Fleet's docs describe extending it "with any MCP server," and the config shape is the same stdio command threadctx already ships.