Studies on object visual working memory have claimed that we can maintain 3-5 objects. However, change detection tasks used in previous work have problems in evaluating feature-bound object representations in working memory. We devised a paradigm called multiple-object permanence tracking (MOPT) for more strict evaluation, where observers are required to identify the type of switch in feature combination between objects during an occlusion period, thus eliminating the use of feature memory or stimulus salience. We showed that capacity of feature-bound representations is more limited than previous estimates. To examine whether this limitation reflects memory retrieval or maintenance, we used a cueing version of MOPT. A flashing cue with 100 % validity was presented on a target object just before or after a feature-switch event. If memory-retrieval is the bottleneck, postcue will facilitate the task performance. A type identification task evaluating feature-bound representations failed to show any benefit of postcue, whereas a simple change detection task possibly reflecting saliency-based representations showed a significant benefit. This suggests that the previously reported capacity of 3-5 objects may reflect saliency-based representations. In contrast, feature-bound representations can be stored only for 1 or 2 objects.