Most CBR systems in operation today are ‘retrieval-only’ in that they do not adapt the solutions of retrieved cases. Adaptation is, in general, a difficult problem that often requires the acquisition and maintenance of a large body of explicit domain knowledge. For certain machine-learning tasks, however, adaptation can be performed successfully using only knowledge contained within the case base itself. One such task is regression (i.e. predicting the value of a numeric variable). This paper presents a knowledge-light regression algorithm in which the knowledge required to solve a query is generated from the differences between pairs of stored cases. Experiments show that this technique performs well relative to standard algorithms on a range of datasets.