My experience with long running sessions is that they lose track of what's going on. The signal to noise ratio is often very poor - some models are particularly verbose and spew a lot of crap. The output of an LLM session is either modification of code or a plan or summary - that has the value, not the session text itself.