Context window vs knowledge cutoff
One is how much the model can look at right now. The other is how recent its training is. They get mixed up constantly, and the fix for each is different.
- → The context window is the text in front of the model right now: prompt, pasted material, and the conversation so far.
- → The knowledge cutoff is the date its training data stops. Nothing after it is in the weights.
- → You get around the cutoff by putting fresh text in the window: paste it, link it, or let the model search.
Two different limits shape what a model can tell you, and people mix them up because both feel like "the model does not know". One is about space: how much text the model can consider at once. The other is about time: how recent the text it learned from is.
A common quiz question puts it well: the context window is what the model can see right now, prompt and pasted text included, and it is separate from the training-data cutoff. True. The rest of this page is why the difference matters.
The context window: what it can see
Everything you send in a conversation, plus the reply being written, has to fit inside the context window, which is measured in tokens. Paste an article and it is in the window. Ask a follow-up ten messages later and the earlier messages are still there, until the window fills and the oldest content drops out. It is the model's working memory, and it is the only thing the model can actually read.
The knowledge cutoff: what it learned
A model is trained on text collected up to some date, then frozen. That date is the knowledge cutoff. Anything published after it, a product launch, an election result, a price change, is simply not in the weights. The model will not always tell you this. Asked about something recent, it may answer from older patterns with full confidence.
Why they are easy to confuse
Both show up as the model not knowing something. If it cannot answer about an event from last month, that is the cutoff. If it forgets a detail you gave it an hour ago in a long chat, that is the window. If it does not know what is in a document you never pasted, that is neither: the text was simply never in front of it.
The fix for each
For the window, keep what matters close: restate key facts in a long conversation, or start a fresh one. For the cutoff, bring the model fresh text: paste the article, paste a link so the page is fetched, or turn on web search and let it look. Anything in the window is available regardless of when it was written, which is why a model with a 2025 cutoff can discuss a page published this morning if you hand it that page.
A related trap: the model knows today's date only if told
Nothing inside the model tracks time. Unless the product tells it the date on every request, it will assume it is still around its cutoff, and search queries or date math will drift. Good products inject the current date. It is a small thing that fixes a surprising number of stale answers.
Think of a researcher who stopped reading the news in a given year but can read anything you put on their desk. The desk is the context window. The year they stopped is the cutoff. Nothing on the desk is too recent for them, and nothing off the desk is available from memory after that year.
Berges AI tells the model today's date on every request, fetches any link you paste so the page lands in the context window, and can search the web under the WebSearch pipeline when a question calls for it. The cutoff still exists in every model we run; those three things are how you work around it.
Try Berges AIRelated concepts
Things people ask
Does a bigger context window fix the knowledge cutoff?
No. A bigger window lets you paste more, which helps only if you paste something. The cutoff is about the training data, and no amount of window space changes what the model learned.
Why does the model think it is an earlier year?
Because its sense of "now" comes from its training data, which ends at the cutoff. If the product does not pass in the current date, the model assumes it is still around when its data ends.
Is pasted text remembered after the conversation?
Only within the conversation, and only while it fits in the window. It does not update the model. The next conversation starts with the model knowing exactly what it knew before.
How do I know a model's cutoff?
The maker usually publishes it, sometimes vaguely. When it matters, ask the model about a well-known recent event and see whether it knows. Do not trust its own claim about its cutoff; models are often wrong about that too.