Continual learning would mean that the data somehow has to be part of the model and the model needs to incrementally adapt to novel inputs. Not just tacked on and backpropagated, but within the network affecting decisions. Current architectures are pretty much dead ends in that aspect.