Checkpoints & Persistence
Without a snapshot, a refund that paused overnight is gone. Checkpoints are how graphs survive deploys, crashes, and humans who come back tomorrow.
A save file in a game. You do not replay the whole dungeon. You load the last checkpoint and keep going.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Overview
Scroll inside the frame to explore · use + / − to zoom up to 200%
Scroll inside the frame to explore · use + / − to zoom up to 200%
Short-term vs long-term
Scroll inside the frame to explore · use + / − to zoom up to 200%
Checkpointer = this thread's working memory. Store = facts across threads (user prefs). Different tools.
Why nodes should be safe to retry
Scroll inside the frame to explore · use + / − to zoom up to 200%
Resume re-runs the node that was in flight. Side effects before the pause can fire twice — make them idempotent.
Key Takeaways
- 1.A checkpointer snapshots state after each super-step so the same thread can resume later. thread_id is the name of that story. Same id = same ticket. New id = a blank run.
- 2.Durable execution is not a second feature — it is this: crash, restart, invoke the same thread, continue. MemorySaver is for local learning. PostgresSaver (or similar) is for production. Pick one at compile time.
- 3.compile(checkpointer=MemorySaver()) then config={'configurable': {'thread_id': 'ticket-4411'}}. Every invoke/stream with that config loads the latest checkpoint first.
- 4.Durable execution in the docs is this mechanism plus retries. Time travel (later) reads the checkpoint history.
- 5.HITL (next) requires a checkpointer or the pause has nowhere to sit.
Learn elsewhere
- →Time Travel & Replay
- →Human-in-the-Loop
Real Example
Scenario
Server dies after classify, before the assistant. Restart. Same thread_id. Graph loads ticket_type and continues at assistant. The user does not start over.
What you would do
One thread_id per conversation or ticket. Never reuse a thread for a different customer. In production, put checkpoints in Postgres, not process memory.
Commands
Commands to Remember
compile(checkpointer=MemorySaver()) # localpip install langgraph-checkpoint-postgres # productionconfig = {configurable: {thread_id: '...'}}Same thread_id resumes, new thread_id starts blank
Cheat Sheet
Quick recap
quick ref- •Checkpointer saves
- •thread_id is the slot
- •Durable = resume
- •Postgres in prod
Common Mistakes
- ✕Invoking without thread_id while a checkpointer is attached — or the opposite, expecting resume with no checkpointer
- ✕MemorySaver in production (RAM dies with the process)
- ✕Non-idempotent refunds inside a node that can be retried