Think of a desk. Whatever is on the desk, the AI can use. Whatever is not on the desk does not exist to it — not what you said an hour ago, not the file you meant to attach. The desk has a size, and this is called the <b>context window</b>.
The model has no memory outside the conversation in front of it. Everything it uses to answer must be in the window at that moment.
This is why it forgets what you told it earlier in a very long chat, and why it cannot use your notes unless you paste them in.
When the desk fills up, the oldest sheet slides onto the floor. The model does not stop. It does not say "chapter 1 is missing". It answers anyway, using what is left, and fills the gap with something plausible.
Most bad AI answers in real work are not clever failures. They are a page on the floor.
Give it the relevant pages, not everything. A 200-page document plus the one table you care about often pushes the table off the desk.
Start a fresh chat for a new task instead of carrying a long history along.
If an answer suddenly gets vaguer, suspect the window before you suspect the model.
It did not say chapter 1 was missing. It said "most likely". And it was wrong.
In our free course you load a four-sheet desk with five chapters, watch chapter 1 fall on the floor, then ask a question and see the answer quietly go wrong.
▶ Play Lemonade Empire free