This study tested the hypothesis that affective content may undermine rather than facilitate working memory (WM) performance. To this end, participants performed a running WM task with positive, negative and neutral words. In typical running memory tasks, participants are presented with lists of unpredictable length and are asked to recall the last three or four items. We found that accuracy with affective words decreased as lists lengthened, whereas list length did not influence recall of neutral words. We interpreted this pattern of results in terms of a limited resource model of WM in which valence represents additional information that needs to be manipulated, especially in the context of difficult trials.