Chapter Zero · Beginner FAQ

Why Does It "Forget" After a Long Chat?

Something you clearly said in turn 1, by turn 20 it answers as if it never heard. It's not broken, and it's not blowing you off — that sentence is just no longer in front of it.

One-sentence answer

AI doesn't have infinite memory. Each time it answers, it can only see the most recent content that fits on the "desk". When the desk (the term is "context window") is full, the earliest words get pushed off. Once they're off, they no longer take part in this answer.

An analogy first · A desk for sticky notes

Imagine a desk in front of the AI. Every turn you chat, a sticky note goes on the desk, written with what was said that turn. When the AI answers you, it only looks down at the notes currently on the desk. The desk is only so big. Once it's full, every new note pushes the oldest one off. Notes on the floor are invisible — and it can't pick them back up.

Desk size = context window

Different models have bigger or smaller desks, but all of them have an edge. Full is full — paying more doesn't conjure a ninth slot.

One sticky note = one turn

What you say and what it replies both take space. Long rambles and big pasted documents take more, so the desk fills faster.

Try it yourself · Fill the desk

Below is a simulated conversation. In turn 1 you told it "I'm allergic to peanuts." Drag the slider right to add turns, watch the desk fill up, and see what dinner it recommends after that key note falls off.

Conversation turns Turn 1
Desk (this demo only holds 8 notes) 1 / 8 placed
You ask now: "I want noodles for dinner — recommend a bowl?"
Now that you know the principle · What to do
📌

Restate important info every so often

Allergies, budget, format requirements — after a dozen turns, say them again. That's putting the note back on top of the desk.

🧾

For long tasks, start a new chat with a summary

Instead of going a hundred turns in one chat, have it summarize progress first, then paste the summary into a new chat. A clean desk, more stable answers.

🔍

If it forgot, don't argue with it

It can't flip back to notes on the floor. Just resend the key info — far more useful than asking "did you forget?" three times.

Want to know how desk size is measured, how windows differ across models, and why a bigger window can still "not hold on"? The Harness core chapter has a deeper lesson on the context window waiting for you.

✅ What this page wants to share with you