That's why I think better maps, and an evolving Tesla network/database of traveled roads, could be a very useful enhancement. Right now, the car is basically experiencing every route as an unfamiliar one, and does no memory imprinting on an individual basis. On a fleet basis, you could say the disengagements will eventually train out some of the mistakes, but traveled-road familiarity is not a primary goal of the training. In one of his last presentations for Tesla, Karpathy did allude to the concept of a virtual whole-earth map being effectively trained into the neural network, but I don't think Tesla has been moving purposefully towards this kind of thing.
In the meantime, I'm hoping the new Chill setting will help in the places where it makes unwanted lane changes - but if there's lane assignment confusion due to ambiguous markings, coupled with the "every drive is my first drive" syndrome, there may not be much difference.