Chapters — How AI Actually Thinks
LLM Fundamentals·Intermediate·6 min read·Sep 5, 2026

Its Short-Term Memory

Why a model can lose track of something said earlier in a long conversation: there's a hard limit on how much text it can actually see at once.

A whiteboard, not a filing cabinet

Imagine trying to solve a problem using only a whiteboard of a fixed size. As you write more, eventually you run out of room, and to keep going, something earlier has to get erased to make space. You don't have a filing cabinet where everything you've ever written is kept and instantly available; you have exactly the whiteboard, and whatever fits on it right now.

That's a fair picture of a model's context window: the maximum amount of text, measured in tokens, that it can actually consider at once when producing its next guess. Everything relevant has to fit on that whiteboard at the same time, your entire conversation so far, any documents you've pasted in, and the response it's currently writing.

What happens when the whiteboard fills up

Once a conversation grows past the context window's size, something has to give. Depending on how the system around the model is built, older messages might get dropped entirely, or quietly summarized to take up less space. Either way, information that no longer fits isn't something the model is choosing to ignore; it genuinely isn't there for it to see anymore, the same way you couldn't reference something you'd already erased from the whiteboard.

This is why a very long conversation can start to feel like it's "forgetting" things you mentioned much earlier. It isn't being careless. The earlier detail may have simply scrolled off the part of the whiteboard the model can currently read.

Whiteboard size trade-offs
A smaller context window
Faster to work with, since there's less to consider on every single guess, but earlier details get pushed out sooner.
A larger context window
Can hold far more, entire documents or long histories, but takes more computation to consider all of it on every guess, and a model doesn't necessarily weigh distant information as carefully as recent information even when it technically fits.

A practical habit this explains

If you've ever been told to "put the important instructions near the end of a long prompt" or to "remind it of something you mentioned a while back," this is why: recent tokens are reliably on the whiteboard and getting full attention. Something mentioned far earlier, especially in a long back-and-forth, is more likely to have faded toward the edge of what still fits, or lost some of the weight it would have gotten if it were closer to the current moment.

Not the same as "it doesn't understand"

Running out of context isn't a sign the model is confused or incapable. It's a hard, physical limit on how much text can be considered at once, the same way you'd struggle to hold an entire novel's worth of detail in mind while writing the next sentence of a book report, no matter how well you understood the book.

Part of a free guide

How AI Actually Thinks

A first-principles guide to what's really happening inside a large language model, built from simple analogies up to the real mechanics.

Browse All Guides →