You already know: the context window is a budget, and summarizing old turns trades detail for room (D7 Context engineering); embeddings find text by meaning (D8 Embeddings); databases give you keys, updates and transactions (D1 Databases); every step of an agent re-sends its history (Agent fundamentals).
Why this matters for Atlas. A model remembers nothing between calls. Everything an agent "knows" about this task, this user or past sessions is something your code chose to put back into the context. Memory is the system that decides what to keep, where to keep it, when to bring it back, and when to throw it away. Done badly, the agent forgets what matters, recalls things that are no longer true, leaks one user's details into another's session, or keeps data it was asked to delete. This lesson builds a working memory system and measures each design choice.
1. A memory hierarchy, like the one in your computer
The question. Where should an agent keep what it knows?
Start from nothing. The same idea as CPU caches and RAM (D1 Operating systems): a small fast tier close to the "processor", backed by large slow tiers.