The introduction of success-history based adaptation in SHADE, a variant of JADE, resulted in a significant advance in the performance of adaptive differential evolution. Many variants of SHADE which use the same success history mechanism have been proposed, but the success history mechanism has remained poorly understood. We revisit use of the success history based adaptation, and show experimentally that the standard approach to sampling from the success history in SHADE may not be as vital to performance as previously assumed. We show that EnJADE, a simple, new variant of JADE which maintains an ensemble of control parameter distribution means, can outperform SHADE on the CEC14 benchmark suite. We also show the effectiveness of the new ensemble-based approach when combined with linear population reduction.