Key takeaways

• A plain LLM is stateless; memory is the external system that gives an agent persistence across steps and sessions.
• Because the context window is limited, you cannot keep everything in the prompt; you need a memory layer that stores and retrieves selectively.
• Having memory can matter more than which model you use.
• Memory is the agent's belief state: if it is wrong, every decision built on it degrades.
• The right mental model is a write-manage-read loop, and the neglected, hardest phase is manage.

Continue to Chapter 2 for the four types of memory.

---