That is how language models and text to speech work. Most language learning models hit a text limit, read over the last inquiry, and figure out based on that and your input, how to respond/ perpetuate the conversation. They can't "remember" what they said in the past. It starts "hallucinating" when the "conversation" diverges too much. You can also waste their tokens by forcing it produce text, images, etc... The whole idea is that you're wasting their investment to scam people.