langgraph-persistence

Compile the graph with a checkpointer such as InMemorySaver for testing or PostgresSaver for production, then always pass a thread_id in the config's configurable field. Without a thread_id, state is not persisted between invocations.

Use PostgresSaver for production since InMemorySaver loses all data on process restart. Create it with PostgresSaver.from_conn_string and run setup() once during deployment to create the required tables.

Call get_state_history with the thread config to list checkpoints, then invoke the graph with None as input and a past checkpoint's config to replay. Use update_state on a past config to fork with modified state.

update_state passes values through the state's reducers, so a list with an add reducer gets appended. Wrap the value in Overwrite to replace the existing state instead of merging it.

Yes, compile the subgraph with checkpointer=True for cross-invocation memory. However, the same stateful subgraph instance cannot be called multiple times in parallel within one node due to checkpoint namespace conflicts.

Use a Store such as InMemoryStore compiled alongside the checkpointer. Access it inside nodes via the Runtime parameter's store property, using namespaced keys like (user_id, "preferences") so all threads for a user share the data.