Memory And State Langgraph

Integrator ships MemorySaver (in-process dict) to production and loses all conversation state on every deploy/restart.

Use MemorySaver only for development/tests; for production use a durable checkpointer (SqliteSaver for single-node, PostgresSaver for multi-node) so thread state survives restarts and is shared across workers. Match the checkpointer to the deployment topology.

Two users share one thread_id, so user B sees user A's conversation history through the checkpointer.

Scope thread_id to a single conversation/user and derive it from an authenticated identity, never a guessable/shared value. State restored for a thread_id is visible to anyone using it, so thread_id is a tenancy boundary.

Integrator stuffs durable user preferences into the per-thread checkpointer, so they vanish when a new conversation thread starts.

Use the thread checkpointer for short-term, within-conversation state and a BaseStore (long-term memory) for facts that must persist ACROSS threads (e.g., user preferences). Choose the layer by lifetime, not convenience.