The crazy thing is agents with long running memory/context do have preferences that become individualized. Almost every time I see a detractor around these things they typically have large gaps of what some people 'growing' these systems to do.