Also, if you then resume your conversation, and the LLM discovers the change, it might revert the "accidental" change.

If your working with an active context, and changes you do then need to be conversed back to the agent, and even then, it might still find it jarring and wrong.