I've set max context to 180k and usually compact around 120k. It's much better to re-read stuff than to have it under-perform when it's over 120k.