A longer context window is not a memory
Context windows have grown from a few pages to the length of several books. It is tempting to read that as a solution to every problem of the form "the model does not know about my data". Just put everything in the prompt.
Sometimes that works. Often it works less well than the numbers suggest.
Fitting is not the same as using
A model can accept a very long input and still pay uneven attention to it. In practice, details buried in the middle of a long document are easier to miss than details near the beginning or the end. When the answer depends on one sentence among thousands, quality becomes sensitive to where that sentence happens to sit.
Every request starts from zero
A context window is also not memory in the everyday sense. Nothing persists between requests unless you send it again. A system that seems to remember previous conversations is storing them somewhere and choosing what to include each time. That choice is the real design problem, and a bigger window only postpones it.
Cost and speed
Long inputs are slower to process and more expensive. Sending a whole manual to answer a question about one setting is like printing the manual every time someone asks. Caching repeated prefixes helps, but the cheapest token is still the one you do not send.
What I do instead
I treat the context window as a desk, not a filing cabinet. Put the instructions and the material needed for the current task on it, keep it tidy, and fetch the rest when needed. Long context is a great tool for tasks that genuinely need a whole document at once, like summarizing a contract or comparing two versions of a report. For everything else, a little selection goes a long way.