Skip to content
Transformers & LLMs

Using LLMs

Context Windows & Memory

9 min+90 XPStep 1/10

The window is the whole memory

An LLM has no memory of past conversations — and even within one conversation it only sees its context window: the tokens currently fed in. GPT-4-class models run ~128k tokens; a few push toward 1M.

That's why a long chat 'forgets' your first message — it doesn't forget, it literally can't see it. Once text scrolls past the window, it ceases to exist for the model.

analogy

An LLM is a genius with a clipboard: brilliant about whatever's on the current page, zero recollection of any page you took away.