We propose a novel feature selection method based on a variable memory Markov (VMM) model. The VMM was originally proposed as a generative model trying to preserve the original source statistics from training data. We extend this technique to simultaneously handle several sources, and further apply a new criterion to prune out nondiscriminative features out of the model. This results in a multiclass discriminative VMM (DVMM), which is highly efficient, scaling linearly with data size. Moreover, we suggest a natural scheme to sort the remaining features based on their discriminative power with respect to the sources at hand. We demonstrate the utility of our method for text and protein classification tasks.