Forgetful Large Language Models: Lessons Learned from Using LLMs in Robot Programming