Study Guide · Chapter 1: State & Memory · 5 min read

Agent Memory, Explained Simply

If you have ever repeated yourself to a chatbot ten minutes into a conversation, you already understand the problem this chapter solves. Most AI models forget everything the moment a conversation window fills up. Agent memory is the engineering discipline that fixes this.

Why AI needs memory

A large language model is brilliant but forgetful. It can only pay attention to a fixed amount of text at once, called the context window — think of it as a small desk that can only hold a few pages.

Anything that does not fit on the desk is gone. So if an agent helps you for three days, everything from day one has fallen off the desk unless someone deliberately saved it. Memory systems are the filing cabinets around the desk.

Three kinds of memory

Working memory is the desk itself: the current goal, constraints, and tool results from the last few seconds. It is fast and small.

Episodic memory is a diary: a durable log of past interactions and outcomes. Semantic memory is an encyclopedia: organized domain facts stored so they can be searched and reused in any future session.

What goes wrong without it

Without tiered memory, an agent asks the same clarifying question twice, contradicts what it told you yesterday, and quietly loses track of multi-step tasks.

Teams often try to fix this by stuffing everything into the context window. That gets expensive fast and lowers quality, because attention spreads too thin. The answer is structure, not volume.

Key Points

  • Models have a limited context window — memory exists to work around that limit.
  • Working, episodic, and semantic memory mirror how people handle the present, the past, and general knowledge.
  • More context is not better context. Tiered, curated memory beats a giant dump of text.
  • Memory design also affects safety: it determines what can be inspected, audited, and forgotten.

All study guides for this chapter: Agent Memory, Explained Simply · How Agent Memory Works Under the Hood · Agent Memory in the Real World