You’re right - poor compression can cause that. But skipping compression altogether is also risky: once context gets too large, models can fail to use it properly even if the needed information is there. So the way to go is to compress without stripping useful context, and that’s what we are doing