There is an interesting contribution in terms of formulating the continual learning setting into a geometry-based selection problem of samples from the experience replay. Reviewers would like to see more experiments on other standard benchmarks, as well as additional discussion about how each approximation step works and the relationship to other existing methods.