LangGraph Checkpointing: Why MemorySaver Doesn't Survive Production

An agent that loses its state every time the process restarts isn't a production agent — it's a demo that got lucky nobody restarted it at a bad moment.

What a checkpointer actually does

A checkpointer is a state-persistence mechanism LangGraph uses to save a graph's current state, store memory between nodes, allow resuming execution after a failure, and maintain conversation continuity across multiple requests. LangGraph saves your graph's state after every individual node execution — not just at the end, but after every single step.

MemorySaver: testing only

Production systems require a durable backend (SqliteSaver for single-process workloads, PostgresSaver or a Redis-backed saver for multi-process deployments) so state survives across the full lifecycle of the process. MemorySaver keeps everything in the process's memory — it works perfectly for local development and tests, and is lost completely on the first restart or when scaling to more than one replica.

PostgreSQL: the standard durable backend

The `langgraph-checkpoint-postgres` library gives you a production-grade checkpointing backend — durable, scalable, and observable. It's the natural choice if your architecture already uses Postgres for the rest of the application, avoiding adding a new piece of infrastructure just for this.

Redis: when latency matters most

RedisSaver is fast, distributed, scalable, and used in real deployments for enterprise systems. Redis offers ultra-fast read/write operations (under 1ms of latency) for saving agent state, with linear scaling for production deployments with growing memory needs — the right choice when checkpoint latency itself is part of the total response latency budget.

How to choose based on your deployment topology

The right checkpointing backend depends on deployment topology: MemorySaver only in tests; PostgresSaver or Redis for any multi-process or containerized deployment. If you already run Postgres, start there — migrating to Redis later, if checkpoint latency becomes a measurable bottleneck, is easier than migrating from MemorySaver to either one under the pressure of a production incident.