What does ChatGPT actually see when you send a message?

The picture most people have

Ask someone how ChatGPT works and you usually get the same story: you're talking to something that listens, remembers what you said earlier, and thinks before it answers. Like a very fast, very well-read person on the other end of the line.

It's a reasonable picture. It's also wrong, and the wrongness explains most of the frustrating things AI chats do.

What actually happens

An AI model has no running memory of your conversation. Between your messages, it holds nothing. When you hit send, the chat application takes everything — the system instructions, every message you wrote, every reply it gave, every file you uploaded — and sends the whole bundle to the model as one big block of text. The model reads that block and produces the most likely continuation. Then it forgets again.

Your twentieth message is not the model “remembering” the first nineteen. It is the model reading all twenty for the first time, every time.

Why the illusion is so strong

Because retransmission is invisible to you, the model behaves exactly as if it remembered. Ask “what did I say my dog's name was?” and it answers correctly — not because it stored the fact, but because the answer was in the text it just reread. The illusion only breaks when the chat gets long, or when the history contains things you no longer want it to see.

What this explains

Why long chats get worse

The model weighs everything in the bundle, not just the part you care about. A long chat is a long bundle. Your question competes with every tangent, error, and abandoned idea you ever typed. This is context pollution, and it's a direct consequence of how messages work.

Why corrections come back

Tell the model it made a mistake and the mistake doesn't leave the bundle. It sits there as a plausible-looking pattern. Twenty messages later, the model re-derives the same wrong answer — because statistically, it's part of the conversation.

Why long chats cost more

You are billed for what gets transmitted. Message forty costs roughly forty messages' worth of input, because message forty contains messages one through thirty-nine. Detours aren't just noise. They're line items.

What this means for how you work

Once you see the mechanic, the strategy writes itself: be deliberate about what's in the bundle. Keep conversations short and on-topic. When a thread has served its purpose, don't let it follow you. When you need a tangent, isolate it instead of mixing it in. That's the entire idea behind branching — and behind Alyph.