Checkpoint

A checkpoint is a snapshot of an agent run’s state, written by the harness so a later process can continue from that point. The snapshot is the thread’s state at a step. It is not the model weights, and it is not the long-term memory of facts the agent has learned about a user.

LangGraph’s persistence page, read November 10, 2026, uses the word for short-term, thread-scoped memory: conversation continuity, human-in-the-loop, time travel, and fault tolerance. You compile the graph with a checkpointer and pass a thread_id. InMemorySaver keeps those snapshots in RAM, so a restart drops them. SqliteSaver and PostgresSaver keep them on disk or in a database. The same page separates a store, which holds key-value data across threads, from the checkpointer, which holds one thread’s graph.

Other harnesses persist different objects. CrewAI Flows use @persist, defaulting to SQLite, to resume or fork a flow’s state. The OpenAI Agents SDK resumes an interrupted run through a session, which stores conversation items. A session replay is not a graph checkpoint. The comparison of those three is LangGraph vs CrewAI vs the OpenAI Agents SDK.

The cycle that produces the next snapshot is the agent loop. A pause for a person is human in the loop. Choosing which node runs next is orchestration.