Cognitive mechanisms underlying the limited transfer effects of working memory (WM) training remain poorly understood. We tested in detail the Strategy Mediation hypothesis, according to which WM training generates task-specific strategies that facilitate performance on the trained task and its untrained variants. This large-scale pre-registered randomized controlled trial (n = 258) used a 4-week adaptive WM training with a single digit n-back task. Strategy use was probed with open-ended strategy reports. We employed a Strategy training group (n = 73) receiving external strategy instruction, a Traditional training group (n = 118) practicing without strategy instruction, and Passive controls (n = 67). Both training groups showed emerging transfer to untrained n-back task variants already at intermediate test after 3 training sessions, extending to all untrained n-back task variants at posttest after 12 training sessions. The Strategy training group outperformed the Traditional training group only at the beginning of training, indicating short-lived strategy manipulation effects. Importantly, in the Traditional training group, strategy evolvement modulated the gains in the trained and untrained n-back tasks, supporting the Strategy Mediation hypothesis. Our results concur with the view of WM training as cognitive skill learning.