Working memory allows information from transient events to persist as active neural representations that can be used for goal-directed behaviors such as decision making and learning. Computational modeling based on neuronal firing patterns in animals suggests that one putative mechanism enabling working memory is periodic reactivation (henceforth termed "replay") of the maintained information coordinated by neural oscillations at theta (4-8 Hz) and gamma (30-80 Hz) frequency. To investigate this possibility, we trained multivariate pattern classifier decoding algorithms on oscillatory brain responses to images depicting natural scenes, recorded with high temporal resolution via magnetoencephalography. These classifiers were applied to brain activity recorded during the subsequent five second maintenance of the scenes. This decoding revealed replay during the entire maintenance interval. Replay was specific to whether an indoor or an outdoor scene was maintained and whether maintenance centered on configural associations of scene elements or just single scene elements. Replay was coordinated by the phase of theta and the amount of theta coordination was correlated with working memory performance. By confirming the predictions of a mechanistic model and linking these to behavioral performance in humans, these findings identify theta-coupled replay as a mechanism of working memory maintenance.