What Is a Context Window? Why AI Forgets — Explained with One Desk Metaphor
You explain your project carefully, the AI responds brilliantly — and twenty messages later it has forgotten the very constraints you started with. That frustrating amnesia has a name: the context window. Understand it once and most of AI's strange memory behavior becomes predictable — and avoidable. No technical background needed; one desk metaphor does most of the work.
The context window is a desk
A context window is the maximum amount of text an AI can consider at once. Picture a desk: everything the model uses to answer you — the conversation so far, pasted documents, your instructions — must physically fit on it. Papers that slide off the desk do not exist for the model. The AI is not "forgetting" the way people do; it simply cannot read what is no longer on the desk.
Why the beginning disappears first
When a long conversation overflows the desk, most services push the oldest papers off first. That is why the setup you gave in message one — the role, the tone, the constraints — is the first thing to vanish in a marathon session. A conversation that suddenly stops making sense is usually a full desk, not a worse model. The fix is unglamorous but reliable: restate the important premises periodically.
Desks are measured in tokens
Window size is counted in tokens — fragments of words, not characters or pages. A rough rule for English is that a token is about three-quarters of a word. Modern models advertise windows from tens of thousands to over a million tokens, and free tiers often get smaller windows than paid ones. The exact limit varies by service and changes often, so check the provider's current documentation rather than trusting any article's numbers — including this one.
A bigger desk is not automatically a smarter reader
Research on long-context models has repeatedly found a "lost in the middle" effect: information buried in the middle of a huge input is recalled less reliably than information at the start or end. Practical consequence: put your most important instruction at the top or bottom of the prompt, never in the middle of page forty. The same principle drives the summarization technique in our guide to summarizing long documents.
Three habits that work with the desk, not against it
- Summarize and restart: in long sessions, ask "summarize what we've agreed so far," then paste that summary into a fresh conversation and continue.
- Repeat what matters: restate key constraints at natural checkpoints instead of trusting message one to survive.
- Paste less, not more: the relevant chapter beats the whole document — for accuracy and speed.
Deciding what goes on the desk, in what order, is itself becoming a named skill — that is exactly what context engineering is about.
Frequently asked questions
Why does my AI remember things from last week, then?
Some services add a separate memory feature that stores selected facts outside the conversation. That is a different mechanism from the context window, and you can usually view and delete what it holds in settings.
How is this different from RAG?
RAG is a librarian: instead of enlarging the desk, it fetches the relevant documents and places them on the desk right when they are needed. Window size limits how much fits; RAG improves what gets chosen.
Can I just paste an entire book?
If it fits the window, yes — but the lost-in-the-middle effect still applies. For anything where accuracy matters, work chapter by chapter and verify claims with our fact-checking routine.
Related on AI Learning Lab: Context Engineering · How to Summarize Long Documents with AI · What Is a Small Language Model?
Comments
Post a Comment