Context window

A context window is the amount of text an AI model can hold in mind at once, measured in tokens (roughly, word fragments). Everything the model reasons from in a single turn, your instructions, the files it is reading, the conversation so far, has to fit inside it. Once you exceed the limit, older content falls out and the model effectively forgets it.

This is the hard constraint behind a lot of AI building frustration. Throw your entire codebase at an agent and you blow the window or bury the signal; the model gets vague and starts making things up. The fix is to treat the window as a scarce resource to curate: feed the model the few things that matter, keep instructions tight, and start fresh when the conversation gets bloated. A focused small context beats a huge noisy one every time.

Say you're drafting a positioning doc in Claude and you paste in twelve transcripts plus your whole website. The early ones quietly drop out of the window and the answer ignores them, so you're better off summarising each first and feeding the summaries. Say you're researching a market in Perplexity: each follow-up question competes for the same window as the sources it already pulled, which is why a long thread starts losing the thread. And say you're loading a quarter of board decks into Notebook LM to ask questions across them, here the larger window is the whole point, but you still get sharper answers by grounding it in the specific documents rather than dumping everything in. Bigger windows help; they are not a licence to be sloppy.