The analogy
Think of a desk you work on. It has a precise surface: it holds, say, twenty sheets. As long as you have few, you see everything at a glance. But you keep laying down new ones, and when you go past twenty, to make room the oldest sheets fall off the opposite edge. You didn't tear them up: they're simply off the desk, and you no longer have them in front of you while you work.
The memory of a conversation with the AI works like this. That surface has a technical name, the context window: it's the amount of text the AI manages to keep "on the desk" while it answers. It includes everything, your questions and its answers. When the conversation stretches past the edge, the first things said fall off, and the AI answers as if it had never heard them. It isn't forgetfulness or bad faith: it's a full desk.
How it really works
The AI doesn't remember in the human sense. With every question of yours, it rereads from the start everything still "on the desk" and answers based on that. If a piece of information has left the window, to the AI it's as if it had never existed: it can't go and retrieve it, because it no longer has it in front of it.
This explains a behavior that's baffling: at the start of the chat you tell it your name, after thirty exchanges you ask it again and it no longer knows it. It didn't "forget" it out of inattention. It's that your name, said fifty messages ago, fell off the edge to make room for everything you wrote afterward.
There's also a subtler effect: as the desk fills up, the AI struggles to give each sheet the right weight. The most recent things, those at the top of the pile, weigh more; those in the middle, even if still on the desk, get less attention. That's why, in a long conversation, an instruction given at the start can fade long before falling off entirely.
What you can do in practice
Knowing the desk has an edge changes the way you work with the AI.
- Repeat the key information when it counts. If after a long chat you ask it something that depends on a detail said at the start, give it back in the same message: "remember I'm vegetarian? suggest me the menu". Don't take for granted that it's still on the desk.
- Keep a summary outside the chat. For long jobs, have a summary of the decisions made given to you now and then and save it in a note. It's your copy of the sheets that fall off the edge.
- Re-paste the useful context when you return to an old topic. Instead of asking "let's go back to that thing from before", re-paste the thing from before. The operating syntax:
Let's pick up a point from before. To be safe I'm giving it to you again, so
you work on the right text:
[re-paste here the information or decision you need]
Now continue from here.
- For important projects, open new chats at clean stages instead of dragging an endless conversation. A fresh chat with a good summary reasons better than a very long one where half the sheets have already fallen off.
A common misunderstanding
The misunderstanding is believing that the AI "gets to know you" as you talk to it, the way a colleague would, learning what you're like by dint of working together. It isn't so. Within a single conversation the AI keeps in mind only what's on the desk, and from one conversation to another, in most cases, it carries nothing: every new chat is born with an empty desk.
From this comes a practical mistake: entrusting the AI with important information counting on it "remembering it for you", and then discovering it has fallen off the window or stayed in yesterday's chat, which is irrelevant. The information that has to last you keep yourself, outside the chat, and bring it back to it when needed. The reliable memory, in this relationship, is yours.
Frequently asked questions
How big is this "desk"?
It varies from one AI to another, and the more recent versions tend to have it larger: some handle very long conversations before losing the initial pieces. But however big, there's always an edge. The practical difference is only after how many pages you start noticing the first lapses, not whether they'll happen.
If I open a new chat, does the AI remember the old one?
In most cases no: every new conversation starts from scratch, an empty desk. Some assistants offer a memory feature that carries some information from one chat to another, but it's limited and has to be turned on: don't take for granted that it's there. If you want continuity, carry the summary into the new chat yourself.
Why does it sometimes seem to remember and sometimes not?
Because it depends on where the information is on the desk. If you said it recently, it's at the top of the pile and it uses it right away. If you said it much earlier but it's still inside the window, it may have faded it and give it little weight. If it has fallen off the edge, it's lost. The same detail, depending on how long ago you gave it, gets different treatment.
So is the AI useless for long jobs?
On the contrary, but it has to be used knowing how it works, and this debunks the myth of an AI that should remember everything on its own. A desk with an edge isn't a flaw to overcome: it's the condition you work under, exactly as you would work yourself, since you don't keep every word of a three-hour meeting in mind either. The difference between those who get frustrated and those who get results in long jobs isn't finding an AI with infinite memory: it's getting into the habit of keeping the important sheets in a safe place and putting them back on the desk when needed. The AI reasons; you guarantee the continuity.