Keep adding text below until the bar turns red.
- The app's hidden instructions~400 tok
- A 40-page PDF you uploadedfell out
1 earlier turn has fallen out. The AI can no longer see them — that's the whole reason it "forgot". A bigger window delays this; it never removes it.
Token counts are estimates (~4 characters per token); real tokenizers differ by model.
That bar is the model's context window: everything it can see at once. When it's full, something has to give.
Anthropic's definition: the context window is "all the text a language model can reference when generating a response, including the response itself." They call it a "working memory".
It isn't what the model learned in training. It's the desk in front of it right now, and it's measured in tokens.
What fills the window on a typical request
Hidden instructions from the app
- Takes space?
- Yes
The whole conversation so far
- Takes space?
- Yes
Files and pasted code
- Takes space?
- Yes
The answer being written
- Takes space?
- Yes, included
How big? It depends on the model. Anthropic's docs describe windows of "up to 1M tokens, depending on the model."
Bigger isn't automatically better. Anthropic: "As token count grows, accuracy and recall degrade, a phenomenon known as context rot."
Try it with the rules off. In the sandbox, fill the window to the brim, then ask a question about the very first paragraph. Then remove half the text and ask again.
Check yourself
0/3 got itSaved on this device only. No account, no streaks.
Next in how the AI thinks: Why the AI "forgot".




