We tested 8‐ and 10‐month‐old infants’ visual working memory (VWM) for object‐location bindings – what is where – with a novel paradigm, Delayed Match Retrieval, that measured infants’ anticipatory gaze responses (using a Tobii T120 eye tracker). In an inversion of Delayed‐Match‐to‐Sample tasks and with inspiration from the game Memory, in test trials, three face‐down virtual ‘cards’ were presented. Two flipped over sequentially (revealing, e.g. a swirl pattern and then a star), and then flipped back face‐down. Next, the third card was flipped to reveal a match (e.g. a star) to one of the previously seen, now face‐down cards. If infants looked to the location where the (now face‐down) matching card had been shown, this was coded as a correct response. To encourage anticipatory looks, infants subsequently received a reward (a brief, engaging animation) presented at that location. Ten‐month‐old infants performed significantly above chance, showing that their VWM could hold object‐location information for the two cards. Overall, 8‐month‐olds’ performance was at chance, but they showed a robust learning trend. These results corroborate previous findings (Kaldy & Leslie, 2005; Oakes, Ross‐Sheehy & Luck, 2006) and point to rapid development of VWM for object‐location bindings. However, compared to previous paradigms that measure passive gaze responses to novelty, this paradigm presents a more challenging, ecologically relevant test of VWM, as it measures the ability to make online predictions and actively localize objects based on VWM. In addition, this paradigm can be readily scaled up to test toddlers or older children without significant modification. Here we introduce a novel eye‐tracking paradigm (Delayed Match Retrieval) to test infants’ visual working memory (VWM) using anticipatory gaze responses. 10‐month‐olds were able to reliably maintain two object‐location bindings over a 1.5‐second delay. Our paradigm can be readily scaled up to test toddlers and older children as well.