This post continues a series on a stateful LangGraph AI agent handling a 15-minute booking workflow. Wrapped with a Streamlit UI, the agent responds to queries, calculates prices, proposes optimized slots, and confirms appointments. To turn this into a proper product, the author replaces in-memory adapters with a dedicated backend.

Previously, the system relied on an in-memory checkpointer for conversation state and a simple Python list for bookings. Since persistence failed upon restarting the demo process, session A could not see bookings created by session B. The agent would even offer slots based on a stale in-memory view, risking double-bookings.

Implementing PostgreSQL fixes these issues by storing bookings and technician information in related tables. This allows multiple frontends to share the same backend safely. While the agent’s graph state still works in memory, persistence and scheduling logic are securely managed by the database, enabling real business operations.