Checkpointing saves snapshots of a graph’s execution state at specific points. In LangGraph, snapshots are taken at each super-step; unlike message-only storage, they preserve the broader graph state. This supports conversational continuity, recovery after failure, and workflows that pause for human input. section 14.1.3
To persist a LangGraph assistant’s state across chat turns:
- Create a thread ID once per conversation and pass it in the RunnableConfig with each graph invocation. section 14.1.4
- Compile the graph with a checkpointer, such as InMemorySaver. For production, use persistent storage so state survives restarts and can be shared across application instances. section 14.1.4
Figure 14.1 Sequence diagram of the travel assistant with checkpoints saved after each node execution, allowing state restoration across turns using a shared thread_id
The sequence diagram shows checkpoints saved after node execution and state restored across turns using a shared thread ID. section 14.1.3
