[dragonwriter]

The only memory an instance of any of the ChatGPT models has is the prompt. The progression of a conversation is adding the past interactions to the prompt. There is nothing to save to. You can do an add-on (say, on the same layer that provides interaction with the user and saving that interaction to the prompt) a facility that does save and retrieval of information on some responses from the LLM rather than sending them to the user, and either train or prompt the LLM to use them, but the recall takes prompt space, since that’s the only place it can be retrieved to for the LLM to use it.

You could compress this with summarization, but eventually your chat history, even with compression, fills up the context space and you don’t have room for new user interaction and responses.