This paper proposes a new approach to discover knowledge about key features together with their degrees of importance in the context of case-based reasoning. A hierarchical memetic algorithm is designed for this purpose to search for the best feature subsets and similarity models at the same time. The objective of the memetic search is to optimize the possibility distributions derived for individual cases in the case library under a leave-one-out procedure. The information about the importance of selected features is revealed from the magnitudes of parameters of the learned similarity model. The effectiveness of the proposed approach has been shown by evaluation results on the benchmark data sets from the UCI repository and in comparisons with other machine learning techniques.