In this paper we propose a new memetic algorithm (MASVM) for fast and efficient selection of a valuable training set for support vector machines (SVMs). This is a crucial step especially in case of large and noisy data sets, since the SVM training has high time and memory complexity. The majority of state-of-the-art methods exploit the data geometry analysis, both in the input and kernel space. Although evolutionary algorithms have been proven to be very efficient for this purpose, they have not been extensively studied so far. Here, we propose a new method employing an adaptive genetic algorithm enhanced by some refinement techniques. The refinements are based on utilizing a pool of the support vectors identified so far at various steps of the algorithm. Extensive experimental study performed on the well-known benchmark, real-world and artificial data sets clearly confirms the efficacy, robustness and convergence capabilities of the proposed approach, and shows that it is competitive compared with other state-of-the-art techniques.