LangGraph 0.2: pluggable checkpointers for durable agents
LangGraph v0.2 split state persistence into dedicated checkpointer libraries (in-memory, SQLite and an open-sourced Postgres version) for session memory, error recovery and human-in-the-loop.
What happened
LangGraph v0.2 was released, focused on checkpointers, the component that saves a graph’s state.
The technical details
From the release notes, there are new dedicated libraries:
langgraph_checkpoint: the base interface (BaseCheckpointSaver), a serialisation protocol, and an in-memoryMemorySaverfor experimentslanggraph_checkpoint_sqlite: SQLite-backed, for local workflowslanggraph_checkpoint_postgres: the optimised Postgres checkpointer from LangGraph Cloud, open-sourced, for production
Imports stay the same (from langgraph.checkpoint.memory import MemorySaver, etc.), and SQLite and Postgres are separate pip installs. The release also includes breaking changes, such as renamed thread_ts and parent_ts.
Why it matters
Persisting state after every step is what lets an agent resume after a crash, keep session memory, or pause for human approval and continue later. The graph in my LangGraph study post had no checkpointer, so its state disappears after each run. A real app with a human-in-the-loop step needs one.
Sources
- LangGraph v0.2.0 release notes — GitHub · langchain-ai/langgraph
This is a summary written from the linked public sources. Figures and numbers are as reported there.