“The recurrent state of the RoboTTT model consists of fast weights, which are parameters updated by gradient descent during both training and inference.”