Why dont machine learning setups ever use over-training as the basis of a memory system?

Elaborate?

Not sure what you are getting at. Somewhat simplified, overlearning to my knowledge is when you learn a unwanted pattern of your input data.

I.e insted of learning how cats look it learns cat faces because data was not balanced between views of cats from all sides. So it wont perform well on a different dataset of cats that does not have this imbalance.

@iiii memorizing the input instead of learning the input.

There has to be some overlooked utility in overfitting

What?

@Wisecrack caching, no?