Stateful graphs with persistent memory

Because graph state shouldn't reset at session end

The Problem

LangGraph is state management done right. Your agent workflows have nodes, edges, and beautiful conditional routing. State flows through your graph like a well-orchestrated symphony.

Within a session.

When the session ends, your graph resets to its initial state. All those carefully accumulated decisions? Gone. The context your agent built up through 15 nodes of processing? Vanished.

You've built a sophisticated state machine that forgets it ever ran.

It's like having a GPS that recalculates from "Starting navigation..." every time you stop for gas.

How Stompy Helps

Stompy adds persistent memory to your LangGraph workflows.

Your graphs gain continuity:

- State snapshots persist: Save important state at key nodes, restore later - Decisions carry forward: What you decided in session 1 informs session 47 - Workflows compound: Each run builds on previous knowledge - Cross-graph memory: Different workflows share the same project context

LangGraph handles the flow. Stompy handles the memory. Your workflows finally have history.

Integration Walkthrough

Install langchain-mcp-adapters

The official LangChain MCP adapters work seamlessly with LangGraph.

pip install langchain-mcp-adapters langgraph

Connect Stompy via SSE transport

Use MultiServerMCPClient with SSE transport and bearer auth.

from langchain_mcp_adapters.client import MultiServerMCPClient
from langgraph.prebuilt import create_react_agent
from langchain_anthropic import ChatAnthropic
import os

# Connect to Stompy via SSE
client = MultiServerMCPClient({
    "stompy": {
        "transport": "sse",
        "url": "https://mcp.stompy.ai/sse",
        "headers": {"Authorization": f"Bearer {os.environ['STOMPY_TOKEN']}"}
    }
})

# Get Stompy's memory tools
tools = await client.get_tools()

# Create agent with persistent memory
agent = create_react_agent(
    ChatAnthropic(model="claude-sonnet-4-20250514"),
    tools
)

Agent saves workflow decisions with lock_context

At key decision points in your graph, save the context. It persists across sessions.

# Session 1: Workflow processes user requirements
result = await agent.ainvoke({
    "messages": [{"role": "user", "content": "Design the notification system"}]
})

# At the "design_complete" node, agent saves:
# lock_context(
#     topic="notification_system_design",
#     content="Architecture: Event-driven with Redis pub/sub.
#              Channels: email, push, in-app. Priority queues
#              for urgent notifications. Rate limiting per user.",
#     priority="always_check",
#     tags="architecture,notifications,design"
# )
# → Creates v1.0

Future workflows continue from decisions

Later sessions pick up from previous decisions. No re-designing settled architecture.

# Session 12: Implementing what was designed
result = await agent.ainvoke({
    "messages": [{"role": "user", "content": "Implement the email notification handler"}]
})

# Agent calls: recall_context("notification_system_design")
# Agent: "Implementing email handler for our event-driven
#   notification system. Using Redis pub/sub as designed,
#   with rate limiting per user. Here's the handler..."

# Architecture decisions persist. Implementation flows.