Long-Term Memory With a Store

Beyond one conversation

Thread state is short-term memory scoped to one conversation. For facts that should persist across conversations (a user's name, language, preferences, past decisions), LangGraph offers a store: a key-value-like memory organised by namespaces (for example ("users", user_id)), usable from nodes and optionally searchable semantically. Decide what to remember (explicit user statements, confirmed facts), write it deliberately, let users view, correct and delete it, and keep sensitive data out unless you have a clear purpose and consent. Retrieve only the few memories relevant to the current task instead of dumping everything into the prompt.

Short-term versus long-term memory

A comparison to guide design decisions.

Short-term (thread state)                Long-term (store)
scope: one conversation (thread_id)      scope: across conversations (namespace, e.g. ("users", id))
content: messages, intermediate values   content: preferences, facts, summaries
saved by: checkpointer after each step   saved by: your nodes, on purpose
lifetime: as long as the thread is kept  lifetime: until deleted; user can view/edit/delete
risk: long threads inflate cost          risk: remembering wrong or sensitive things

Quick check: Which belongs in long-term memory?

• A node's local variable
• The intermediate tool result of one step
• A user's preferred language across conversations
• The recursion counter