An AI model remembers nothing about your conversation. Each time you send a message, the whole transcript so far is handed to it fresh, it reads all of it, and it writes the next reply. The context window is the size of the desk that transcript has to fit on. Once the desk is full, something slides off, and it is usually the oldest part: the constraint you set in message one.
The short answer: a context window is the total text a model can hold in view at once, counting your instructions, every message, any uploaded files and the reply it is writing. Past the limit the oldest parts drop out, and quality starts slipping well before the limit arrives.
Windows are counted in tokens. One token is about four characters of English, so 1,000 tokens is roughly 750 words. Code runs denser, and languages the tokenizer covers less well can take two or three times as many tokens for the same meaning.
| Context window | Roughly holds |
|---|---|
| 8,000 tokens | A long article |
| 32,000 tokens | A short research paper with notes |
| 128,000 tokens | A 300 page book |
| 200,000 tokens | A dense technical manual, or a mid size codebase |
| 1,000,000 tokens | Several books, or a year of meeting transcripts |
The limit applies to the running total of the conversation, the model's own replies included, and never to one message on its own.
On every turn the model rereads the system prompt (often thousands of tokens you never see), any stored memory, every earlier message in full, every file, every search result, and the reply it is producing. A 40 page PDF alone is roughly 20,000 to 30,000 tokens. Eight short messages from you can sit beside eight long answers, two documents and three searches, and your typing ends up as maybe ten percent of the total. Memory features do not help. They save facts between chats and inject them into new ones, which spends window instead of adding any.
Quality fades gradually, with no cliff at the limit. In the Lost in the Middle experiments (Liu et al., 2023), moving the key document from the edges of the input to the middle cost models roughly 20 points of accuracy. Finding one fact in a long input is a solved problem. Combining four scattered facts is still hard, and that is exactly what people buy big windows for. Models that share the same million token window score anywhere from 9 to 51 on the Intelligence Index.
A million token window guarantees the model can accept your document. It guarantees nothing about how well the model reads it.
The habits that fix most "it forgot" moments are cheap. Start a new chat when the topic changes. Paste the long material first and ask your question after it. Restate the rules that matter after about fifteen exchanges. When a session runs long, ask for a written handoff (current state, decisions, open questions, constraints) and carry it into a fresh chat. Rolling summaries are lossy on specifics: "never use the word synergy" can shrink to "the user has style preferences". The context window comparison lists the current size for each model.
The full explainer covers what happens when you run out, seven tactics in order of payoff, and how much window each kind of work needs: What is a context window? The AI limit that explains everything on whizi.io.
Whizi puts 280+ models in one chat, so a long document can go to a large window model and the writing to another without changing subscriptions. The trial is $0.99.