Skip to main content
One of the most common mistakes in agent design is treating “State” and “Memory” as the same thing. In StateBase, they serve two distinct purposes.

​
1. State (Active Logic)

Think of this as the “RAM” of your agent. State is a structured JSON object representing the current condition of the agent. It is what the agent uses to make its immediate next decision.
  • Structure: Strongly typed schemas (e.g., {"status": "onboarding", "step": 2}).
  • Behavior: Updated frequently, often every turn.
  • Goal: Deterministic logic flow.
  • Example: A shopping assistant’s current cart_items list.

​
2. Memory (Long-term Facts)

Think of this as the “Hard Drive” of your agent. Memory consists of unstructured or semi-structured facts extracted from previous interactions. It is retrieved semantically when relevant.
  • Structure: Natural language or key-value pairs ("User is allergic to peanuts").
  • Behavior: Static until reinforced or corrected.
  • Goal: Personality and long-term personalization.
  • Example: The fact that a user lives in Seattle and prefers python.

​
The Synthesis: Context

When you call sb.sessions.get_context(), StateBase merges these two:
  1. It retrieves the Current State (Cart items: 3).
  2. It fetches Relevant Memories (User usually buys gluten-free).
The Result: A perfectly primed prompt that is both logically accurate and personally relevant.