LangGraph state vs memory

Short answer

State is the data a graph run reads and updates. A checkpoint is a saved state snapshot. A thread groups the checkpoints for one ongoing conversation or job. A store can hold information that needs to be shared across threads.¹

Tiny example

For a support conversation:

state      = current messages, selected account, proposed action
checkpoint = state after the user confirms the account
thread     = support conversation 9f21
store      = the user's saved communication preference

Calling all four “memory” hides important lifecycle and security decisions.

Strong follow-up

Explain that persistence requires product choices: retention, encryption, tenant isolation, deletion, and what may be used in future threads. Checkpointing execution does not grant permission to retain everything indefinitely.

A message list is one possible part of short-term application memory, not proof that a model autonomously remembers. For the complete lifecycle comparison, read message history is not durable memory.

Footnotes

1. LangGraph persistence, official documentation. ↩