Custom gradient with state

Question

I\'m trying to implement this gradient clipping paper in tensorflow, which entails storing the history of gradient norms.

I assume I need to do this using the t