LangGraph gives you durable workflow state. A memory layer gives your agent durable user and session recall. If you are building production graphs, you usually want both.
LangGraph is strong at orchestrating multi-step workflows and preserving graph state between steps. That solves a real problem: the agent can continue a run without losing the execution path.
But graph state is not the same thing as user memory. Checkpoints tell you what the workflow was doing, not necessarily what the user prefers or what should persist long after the run ends.
A memory layer sits beside the workflow and stores durable facts, instructions, and preferences. Before a graph step calls the model, it can retrieve the most relevant memories for that user and combine them with the current graph state.
That keeps the graph deterministic while still letting the agent behave like it remembers the relationship.
Start by writing memory only at key points: after onboarding, preference capture, support resolution, or user corrections. Then retrieve memory at the graph entry point and before the most context-sensitive nodes.
This avoids bloating every step while still improving continuity where users feel it most.
Use workflow state for execution and a memory layer for continuity across runs.
These guides reinforce the memory, context, and benchmark cluster this article belongs to.