This is conversational momentum due to the autoregressive nature of models. Each token is drawn from a probability distribution conditioned on the preceding tokens. The best method I know of working around this is to request and curate a markdown export package from a conversation that is used to prime a new conversation in a clean context.