It is hypothesized that working memory (WM) mediates the effectiveness of corrective feedback as a cognitive factor. Yet little is known about the mediating process itself. The present study investigates whether corrective recasts (CR) facilitate L2 learning and examines whether recast-driven L2 learning can be predicted by different aspects of WM. To this end, a pretest-immediate posttest- delayed posttest design was employed with three classes of first-year English majors from a Chinese university as test control, task control and CR groups. The effect of CR was evaluated using oral production (OP) tests and WM was measured by means of nonword span, digit span and listening span tests. The results showed that the CR group significantly improved its accuracy in the use of the final "s" of the third person singular from pretest to posttests and outperformed the test control group on the immediate posttest. There were differential relationships between the efficacy of CR and different components of WM. While nonword span predicted immediate and delayed OP performance, listening span predicted only the immediate language development and digit span seemed less relevant to L2 learning.