MODULE 01 — FOUNDATIONS · 04 AI Concepts

The context window

The model has no memory of your conversation. It has a desk — what's technically called the context window — and everything on it gets re-read before every word, right up to the desk's own edge.

fallen off
On the desk0 Desk holds2,000
DESK SIZE

Block widths are drawn to scale against the desk. Very short messages are widened slightly so you can still see them.

The problem it explains

It doesn’t remember. It re-reads.

People assume the model carries the conversation with it. Nothing carries over — the entire transcript is handed back in, from the top, before each new word.

A PERSON carried memory + newest sentence → carry on A MODEL cut edge of the window re-read in full one word out then the whole desk is read again, from the top
Everything left of the red line was said, but is no longer handed over — so as far as the model is concerned it was never said.

How it works

Three things worth knowing

turn 1 turn 2 turn 3 the whole transcript, re-sent every time
Rebuilt every turnYour app quietly re-sends the lot. That is the only reason chat has continuity.
desk length × words out = work
Everything on it costsLong chats get slower and pricier for exactly this reason.
gone edge no warning
The edge is silentIt answers as though the fallen-off part was never said.

See the distinction

Learned versus on the desk

These are the two different things people fuse into the word “knows”.

the model’s numbers frozen at training “no, it’s 12” identical for every user, every chat
Baked inSet once, unchangeable by talking, the same for everybody.
the desk refilled every chat · swept when you close it
On the deskDifferent every chat, beats what it learned, and gone when you close the tab.

So what

Almost every trick later in this series — looking documents up, giving an agent a memory — is a way of deciding what deserves space on the desk.

MODULE RECAP

Before you move on

Click an idea below, then click the card it matches.

Foundations, connected: language becomes numbers, a machine predicts the next one, and everything it can see lives in one bounded space.

Continue to Module 2 →