The amount of visual information that can be stored in working memory is inherently limited, and the nature of this limitation has been a subject of intense debate. The debate has relied on tasks and models that assume visual items are independently encoded in working memory. Here we propose an alternative to this assumption: similar features are jointly encoded through a 'chunking' process to optimize performance on visual working memory tasks. We show that such chunking can: 1) facilitate performance improvements for abstract capacity-limited systems, 2) be optimized through reinforcement, 3) be implemented by neural network center-surround dynamics, and 4) increase effective storage capacity at the expense of recall precision. Human subjects performing a delayed report working memory task show evidence of the performance advantages, trial-to-trial behavioral adjustments, precision detriments, and inter-item dependencies predicted by optimization of task performance though chunking. Furthermore, by applying similar analyses to previously published datasets, we show that markers of chunking behavior are robust and increase with memory load. Taken together, our results support a more nuanced view of visual working memory capacity limitations: tradeoff between memory precision and memory quantity through chunking leads to capacity limitations that include both discrete (item limit) and continuous (precision limit) aspects.