Articles
What is a context window, and why AI forgets you
You explained everything carefully at the start, and twenty messages later the model knows none of it. That is not a bug — it is how the thing works.
A familiar situation: you start a conversation, explain the context carefully, get good answers. Twenty messages later the model starts ignoring things you established at the beginning.
You are not unlucky. You reached the edge of the context window.
What it is#
A model does not have memory in the sense that you do. It has a window — a fixed number of tokens that can sit in front of it at once.
Everything has to fit inside: the system prompt, your question, the model's own previous answers, every uploaded file. When the conversation outgrows the limit, the beginning slides out. As far as the model is concerned, it no longer exists.
How big#
On current models, from 128,000 tokens to several million. That sounds like a lot, and it is — 128,000 tokens is roughly 300 pages of English text.
But about a third less Estonian fits in the same window, because the same text costs more tokens. Upload a couple of long documents and a meaningful share of the window is already spoken for.
Three effects you will notice#
The AI forgets. The most obvious one. Early instructions disappear.
Answers get more generic. With the specific context out of the window, the model falls back on general knowledge.
It repeats itself. If it can no longer see what it already said, it says it again.
What to do#
Start new conversations more often. Each new topic deserves a fresh window. Continuing an old conversation carries the whole history along, and costs tokens for it.
Restate what matters. If something is critical, say it again rather than referring back to "as I said earlier".
Summarise midway. "Summarise this conversation in five points" — then start a new conversation with that summary. You are compressing the context.
Upload only what is needed. Do not attach a hundred-page report when three pages are the relevant part.
Why not just make the window bigger#
They do, constantly. But a window costs: compute grows faster than the window does, and in a very long window models tend to use material in the middle less well than material at the start and end.
A bigger window is neither free nor a magic fix. Choosing context carefully stays useful even when the window is large.
