But true CL is the ability to learn out of distribution information on the fly.
The only true solution I know to continual learning is to completely retrain the model from scratch with every new example you encounter. That technically is achievable now but it also is effectively useless.