LangGraph and LangChain

A LangGraph workflow can read what a previous session learned, then leave a note for the next session. The runnable example uses LangChain’s built-in MCP adapter and Claude-Mem’s installed MCP server. It adds no service, database, model provider, or global configuration.

What this connection does

The graph has two nodes. recall searches one project, checks a caller-selected observation ID, reads its timeline, and fetches that observation. remember saves the handoff text the caller supplied. The returned observation is evidence; it cannot select another project or authorize another action.

Capability	This example
Recall	`search`, `timeline`, and `get_observations`, scoped to the supplied project
Explicit save	`save_memory` after recall succeeds
Automatic conversation or tool capture	Not provided
Automatic context injection	Not provided; the recalled text is returned to your application
Model calls	None; the graph runs deterministically without a model key
CMEM Pro	Not required; this guide uses the local worker runtime

Tested on Linux with Python 3.12.14, LangGraph 1.2.20, LangChain 1.4.10, FastMCP 4.1.8, langchain-core 1.7.7, MCP Python SDK 2.3.0, and Claude-Mem 15.16.1. macOS and Windows were not verified. These are the versions tested, not a promise about every earlier release. LangChain marks langchain.mcp as beta and emits a warning when it is imported; its API can change. The example uses this maintained API, which replaces the archived langchain-mcp-adapters package.

Connect

You need an installed Claude-Mem worker and MCP server, Node.js, and Python 3.11 or later. Run these commands from a Claude-Mem repository checkout:

python -m venv .venv-langgraph
. .venv-langgraph/bin/activate
python -m pip install -r examples/langgraph/requirements.txt
export CLAUDE_MEM_MCP_SERVER="/absolute/path/to/claude-mem/plugin/scripts/mcp-server.cjs"

Use the installed MCP script you already trust, or the shipped script in this checkout. The path is passed as one process argument and may contain spaces. Set CLAUDE_MEM_DATA_DIR and CLAUDE_MEM_WORKER_PORT only if your existing worker uses nondefault values. Keep them consistent with the worker you intend to use. The example selects the local worker runtime and disables worker autostart for its MCP subprocess; it does not alter saved settings or the worker’s account. Check or start the worker using the existing commands from the checkout:

npm run worker:status
npm run worker:start

Reinstalling the Python requirements does not register hooks or modify an MCP configuration file. There is no integration installation state to duplicate.

Run a handoff

Choose an observation already saved in the project. For a reproducible example, you can explicitly save this harmless setup note to the worker first:

curl --fail-with-body "http://127.0.0.1:37777/api/memory/save" \
  -H 'Content-Type: application/json' \
  -d '{"project":"shop-api","title":"Scheduler timezone","text":"The quokkaharbor scheduler writes dates in UTC."}'

Use the returned observation id below. Replace the port if yours differs. The project name is the same exact key used by the worker and viewer; use the key for your actual checkout when sharing memory with a coding agent.

python examples/langgraph/memory_workflow.py \
  --project shop-api \
  --query quokkaharbor \
  --observation-id 123 \
  --run-id scheduler-handoff-1 \
  --note "Next run: keep the scheduler export in UTC."

The command prints the recalled observation, the worker’s save confirmation, and the saved ID. It exits nonzero if the selected ID is absent from this project’s first 20 search results, recall fails, or the worker does not confirm the save. Choose a narrower query when the selected result is outside that page. The core connection uses LangChain’s built-in adapter and an explicit stdio transport:

from fastmcp import Client
from fastmcp.client.transports import StdioTransport
from langchain.mcp import MCPAdapter

client = Client(
    StdioTransport(command="node", args=[str(script)], env=env, keep_alive=False),
    timeout=10,
    init_timeout=10,
)
async with MCPAdapter(client) as adapter:
    tools = await adapter.list_tools()
    # Build and run the StateGraph while this adapter is open.

The complete example calls each tool with a LangChain ToolCall envelope and checks the returned ToolMessage.status. An MCP execution error cannot be mistaken for a successful result. It filters discovery to the four memory tools it uses and supplies project for every call. Reads also receive the explicit single-item projects=json.dumps([project]) list, so the MCP server cannot widen the selected worktree scope to its parent checkout. Saves receive only project. The project list is encoded as a JSON string so the GET query path preserves literal commas and JSON-looking keys. Nonempty project names retain their exact outer whitespace. Native checks seed separate split/trimmed-key projects and confirm they stay out of search, timeline, and recalled details.

Identity and repeated runs

LangGraph state contains the project, selected observation ID, and caller’s run_id. Claude-Mem stores explicit notes in its existing manual session for that project and langgraph platform. This example does not create or impersonate captured LangGraph sessions. The note records integration: "langgraph" and runId as metadata. A fingerprint of the integration, project, run ID, and note text uses the worker’s existing sourceFingerprint deduplication. Deliberately running the same handoff again returns the same observation ID; changing the note produces a new save. There are no automatic write retries. If a request times out after a possible save, inspect the viewer or rerun with the same run ID and text.

Verification

The native integration checks start the shipped worker with a temporary SQLite data directory, disable optional capture, sync, and telemetry for that fixture, and call the installed LangGraph and MCP adapter. They need Bun and Node.js:

(
  unset CLAUDE_MEM_CLOUD_SYNC_TOKEN CLAUDE_MEM_CLOUD_SYNC_USER_ID CLAUDE_MEM_CLOUD_SYNC_HUB_URL
  export CLAUDE_MEM_MEDIA_CAPTURE_ENABLED=false
  python examples/langgraph/test_workflow.py
)

This verification uses fresh worker settings and a restricted child-process environment. It imports the shipped WorkerService and calls start() directly, so the CLI startup path does not run host hook-configuration repairs. The subshell also clears inherited cloud-sync credentials; it does not change your normal worker’s settings or saved memory. The verified run passed eight checks: recall and save with repeat deduplication, foreign-project ID rejection, blank-project rejection, a stopped-worker failure, and a bounded slow-worker response. The successful run recalled observation 1, saved handoff 3, and returned 3 again on the deliberate repeat. The data folder contained spaces. The timeout case uses a controlled slow HTTP endpoint; success and project isolation use the real bundled worker and SQLite. Two checks create an actual Git repository and linked worktree, seed matching parent/child observations, and run the native graph from the worktree. The parent stays out of the child’s index and timeline, and a parent ID is rejected without adding any SQLite observation.

Data and failure handling

The project, explicit query, selected ID, note, and provenance metadata are sent to the local worker. The recalled text is returned to the Python application. This example sends nothing to a model. Existing worker settings still govern any enabled cloud sync or exports of saved notes. If you add a model or framework tracing to your own application, review where that application sends recalled text. Do not put credentials in a note or use this example as an automatic capture hook. Project filters select memory within the current local user’s worker; they are not a multiuser authentication boundary. The example uses the current MCP manual-save contract and does not import conversations or change privacy settings. Each MCP request has a 10-second timeout. An 80-second outer bound also covers session setup and the complete graph. The command reports a short failure without echoing tool arguments or transport exceptions. Check the worker status, the script path, and the selected project and ID before retrying. A missing worker is not started behind your back.

Disconnect

Stop running the script or remove the call to run_workflow from your application. The context manager closes its MCP subprocess when the graph ends. You can delete the dedicated Python virtual environment. No hooks, user configuration, or saved memory are removed.

Get Started

Cursor Integration

Grok Bot Integration

Antigravity CLI Integration

Best Practices

Configuration & Development

Hosted Server

Architecture

LangGraph and LangChain

What this connection does

Connect

Run a handoff

Identity and repeated runs

Verification

Data and failure handling

Disconnect

Official references

​What this connection does

​Connect

​Run a handoff

​Identity and repeated runs

​Verification

​Data and failure handling

​Disconnect

​Official references

What this connection does

Connect

Run a handoff

Identity and repeated runs

Verification

Data and failure handling

Disconnect

Official references