Distribution shifts between training and test data are all but inevitable over the lifecycle of a deployed model and lead to performance decay. Adapting the model can hopefully mitigate this drop in performance. Yet, adaptation is challenging since it must be unsupervised: we usually do not have access to any labeled data at test time. In this paper, we propose a probabilistic state-space model that can adapt a deployed model subjected to distribution drift. Our model learns the dynamics induced by distribution shifts on the last set of hidden features. Without requiring labels, we infer time-evolving class prototypes that serve as a dynamic classification head. Moreover, our approach is lightweight, modifying only the model's last linear layer. In experiments on real-world distribution shifts and synthetic corruptions, we demonstrate that our approach performs competitively with methods that require back-propagation and access to the model backbone. Our model especially excels in the case of small test batches - the most difficult setting.