Developmental improvements in working memory (WM) maintenance predict many real-world outcomes, including educational attainment. It is thus critical to understand which WM mechanisms support these observed behavioral improvements, and how WM maintenance strategies might change through development. One challenge is that specific WM neural mechanisms cannot easily be measured behaviorally, especially in a child population. However, new multivariate decoding techniques have been designed, primarily in adult populations, that can sensitively decode the contents of working memory. The goal of this study was to deploy multivariate decoding techniques to decode the contents of WM in children. We created a simple computerized WM game for children, in which children maintained different categories of information (visual, spatial or verbal). We collected electroencephalography (EEG) data from 20 children (7-12-year-olds) while they played the game. Using Multivariate Pattern Analysis (MVPA) on children's EEG signals, we reliably decoded the category of the maintained information during the sensory and maintenance period. In a set of exploratory reliability and validity analyses, we examined the robustness of these results when trained on less data, and how these patterns generalized within individuals throughout the testing session. Furthermore, these results matched theory-based predictions of WM across individuals and across ages. As the first study of its kind, our proof-of-concept provides a direct and age-appropriate potential alternative to exclusively behavioral WM maintenance measures in children. Our study demonstrates the utility of MVPA to directly measure and track the spontaneously-generated representational content of children's WM.