Humans and recurrent neural networks both allocate working memory resources to maximize utility, shaped by natural stimulus statistics and learned stimulus-reward association.