langgraph-persistence

Compile your graph with a checkpointer and always pass a thread_id in the config. Use InMemorySaver for testing or PostgresSaver for production, then invoke with {"configurable": {"thread_id": "session-1"}} so state persists across calls.

Use a Store such as InMemoryStore compiled alongside the checkpointer. Save data with store.put under a user-scoped namespace, and access it inside nodes via the Runtime parameter's runtime.store, making it available across all thread_ids.

InMemorySaver suits testing and development but loses all data on process restart. PostgresSaver is production-ready, persists checkpoints in PostgreSQL, and requires a one-time setup() call to create tables.

The most common cause is missing thread_id in the invocation config, which prevents state persistence. Verify you compiled the graph with a checkpointer and pass {"configurable": {"thread_id": "..."}} on every invoke call.

No. Subgraphs compiled with checkpointer=True cannot be called multiple times within one node because calls write to the same checkpoint namespace and conflict. Wrap each subgraph in its own StateGraph with a unique node name for namespace isolation.

update_state passes values through the state's reducers, so a list with an add reducer gets appended rather than replaced. Wrap the value in Overwrite from langgraph.types to replace the state value directly.