This paper makes revisits the long-term memory problem (switching within a small set of experts) in the full info and bandit cases, with specific focus on obtaining improved results in the stochastic case. The reviewers are all appreciative of the new black-box reduction proposed in this paper, and the way in which it is applied to the sparse bandit problem. Of course it would be nice to have lower bounds and experiments... Yet the consensus is that the paper makes a sizable contribution as-is. During the review the question was raised whether this arguably niche problem would find sufficient interest at NeurIPS. Past instances of the conference did host papers discussing aspects of this probem, see e.g. references [2], [13] and [34]. It appears that worry about this is unfounded.