I really don’t think the long term memory will be too much of an issue. I believe you will get a branched version of the model catered to a specific user that’s learning all the time, similar to how instuctGPT works but new data is stored and embedded many times per day