Hello,
I believe that ChatGPT should manage a server-side JSON model of the user, rather than saving random summarized excerpts of conversations.
Perhaps it could add to a memory by thinking something like <<<<user.family.sister.name=“Bertha”>>>> and its memory model would be updated.
The JSON memory model of the user would then be provided to it with every request—possibly pruned by a smaller LLM for relevance to the present conversation.
I’m suggesting this because I believe it likely more closely resembles how our brains remember facts: by building models of the things around us. The current approach could be compared to forcing the model to memorize the entire summarized memory except, word for word.
Best regards,
Greg