Lesson 18 — Short-term memory: the conversation window

Keeping the history small

Models read text in tokens — small pieces of words (roughly 3–4 English characters each). Every model has a context window: the most tokens it can read in one call. Providers also charge per token, so a long history is slower and costs more on every call.

Short-term memory is simply the message list. To keep it small, agents use a window: keep the system message (the rules must never be lost) plus only the last few messages.

rest[-4:]   # the last 4 items of a list

Watch out: build a new list — do not delete from the original history; you may want it later. And rest[-0:] is the whole list, not an empty one.

Why it matters: long chats break agents that never trim — they hit the context limit or get expensive.