I am adjacent to personal LLMs at work. They've been trying to make them work for a number of years. LLMs make it simpler to query the data, but its not quite

I am adjacent to personal LLMs at work. They've been trying to make them work for a number of years. LLMs make it simpler to query the data, but its not quite there yet for useful work. For example if you have an assistant and you talk off and on about booking for $band, they will know that you like music, and that genre. They can alter the context of replies to suit. With LLMs, the link between shortterm and long term recall is broken. Now, this might be partially solvable by having an intermediary LLM that search the interaction database for context and feeds it into the main interaction LLM. But I'm not convinced that its a long term or viable solution. the LLM is not building a personal insight into your responses, its just having one shot context injected.