We investigated whether clustering based on feature similarity improves the representational quality of visual working memory (VWM). We hypothesized that similar items are organized into clusters, and their recall precision increases with fewer clusters because of reduced memory load. In a series of 6 experiments, participants remembered orientations or colors of several stimuli and estimated the orientation (color) of cued item(s). We measured recall bias to identify whether items formed cluster(s) and measured recall precision to determine the effect of clustering on the representational quality of VWM. In Experiments 1 and 2, orientation similarity was manipulated to partition stimuli into 1, 2, or 3 clusters. In Experiment 3, we varied both the number of stimuli and their similarities such that 5 items were summarized into a smaller number of clusters than 3 items. We consistently found that similar items formed a cluster, and that the precision of the individual items increased with fewer clusters regardless of the number of items. We also observed the same clustering effects using color stimuli when participants were to remember items' colors (Experiment 4). However, a task-irrelevant feature was not potent enough to cluster items and did not increase the precision (Experiment 5). In Experiment 6, we varied item similarity and found that response errors were correlated within the same cluster but not across different clusters. Taken together, these results suggest that clusters formed by similar items can impact the representation of VWM; thus, acting as one of representational units of VWM.