There are many applications where the world being interpreted via clusters can change. We present a method that discovers new clusters and describes the changes. The method works by constraining existing prototypes while penalizing changes in a variable, total number of clusters. This results in a clustering that is comparable to the old yet still flexible enough to learn new behaviors. Moreover, the results are highly interpretable. The paper offers two main contributions. One, we present a framework that distinguishes different types of change of interest. Two, we present a new cluster-based change description algorithm and test, both of which are applicable to multiple underlying clusterers.