One form of context rot is what I call self-reinforced structure. When you accept a long-form model answer, you signal that this structure is acceptable, and so it tries to generate subsequent responses in a similar way. It can be destructive for any long-form creative work.

The only real defense is ensuring that the history the model receives doesn’t contain such replies. So it should be either prevented early or later fixed by providing a summary of the previous conversation instead of the actual history.