This experiment addresses the question of what makes a working memory measure a good predictor of higher-level abilities. Verbal and visuospatial processing episodes were interleaved with distinct verbal and visuospatial storage episodes to form four complex span tasks. Although these measures were reliable predictors of reading and mathematics ability in children, they were no more predictive of these abilities than corresponding simple span tasks involving storage alone. However, when individual differences in storage ability and processing capacity were controlled for, residual variance in complex span performance was related to academic ability in some cases. These findings indicate that complex span tasks are multiply determined, and that differences in task structure can dramatically influence the relative importance of these multiple constraints and the predictive power of a complex span measure.