This work considers bases for working memory for non-verbal sounds. Specifically we address whether sounds are represented as integrated objects or individual features in auditory working memory and whether the representational format influences WM capacity. The experiments used sounds in which two different stimulus features, spectral passband and temporal amplitude modulation rate, could be combined to produce different auditory objects. Participants had to memorize sequences of auditory objects of variable length (1-4 items). They either maintained sequences of whole objects or sequences of individual features until recall for one of the items was tested. Memory recall was more accurate when the objects had to be maintained as a whole compared to the individual features alone. This is due to interference between features of the same object. Additionally a feature extraction cost was associated with maintenance and recall of individual features, when extracted from bound object representations. An interpretation of our findings is that, at some stage of processing, sounds might be stored as objects in WM with features bound into coherent wholes. The results have implications for feature-integration theory in the context of WM in the auditory system.