Course contentsShow
← Machine Learning and Deep Learning
Lesson 2060 of 3,53844. AI Agents: FoundationsFree lesson

Agent State and Memory

Learn how agents maintain internal state to track conversation history, observations, and intermediate results.

Agent State and Memory

What You'll Learn

You'll learn how AI agents maintain internal state to track conversation history, observations, and intermediate results across multiple steps.

Why Agents Need Memory

Unlike a simple LLM call that processes one input and returns one output, an agent operates over multiple turns. It needs to remember:

  • What it's already tried
  • What information it has gathered
  • The conversation history with the user
  • Intermediate reasoning steps
  • Which tools it has called and their results

Without memory, an agent would repeat itself, forget context, and fail to build on previous observations — like a detective who forgets every clue immediately after finding it.

Types of Agent State

Short-term (working) memory stores the current episode:

  • The original user query
  • Retrieved documents from this session
  • Tool outputs from recent actions
  • The chain of reasoning so far

Long-term memory persists across sessions:

  • User preferences learned over time
  • Summaries of past conversations
  • Historical success/failure patterns

Think of it like a doctor's appointment: short-term memory is everything discussed today, while long-term memory is your medical history file.

Implementation Patterns

Agent state is typically stored as a structured object or dictionary that gets updated after each action. For example:

  • conversation_history: list of messages
  • observations: facts gathered this session
  • tools_used: which tools were called
  • current_goal: what the agent is trying to accomplish

The agent reads this state before deciding its next action and writes to it after executing each step.

Key Takeaway: Agent state and memory enable multi-step reasoning by preserving context, observations, and history across the perception-action loop — transforming stateless LLM calls into stateful, goal-directed behavior.