[zbentley]

I guess, but like ... long context/context rot/contextual memory retrieval are still very unsolved problems. If an AI's getting input from multiple family members for years, I have a lot of trouble believing that current LLM technology will ever be able to synthesize ancient information into current conversations.

I think that's a structural defect in AI, not one of those "will be improved by the next batch of models/hardware" things. The distant-memory systems that do exist are fundamentally modeled on topic-based/feature-based retrieval. Assistant-type family AIs will probably work OK with that ("what foods do I hate" is probably reliably retrievable), but interpersonal relationships seem to be very different: rather than retrieving based on feature/fact proximity, a conversation in a long relationship has a very gestalt-ish behavior in that it's informed by lots and lots of distant memories at once. That sounds more like actual training, rather than context/memory/post-training, to me, and current AIs are not set up for that.