Visual working memory (WM) holds a limited amount of information for immediate use. Recent work has highlighted that the information in visual WM can be represented in at least two ways: either as features bound into unified object-based representations or as flat sets in a feature-based representation. WM representations may provide a crucial substrate for reinforcement learning. If this is the case, then whether learning operates on objects or features should be determined by the type of representations in WM. To test this hypothesis, we coupled a reinforcement learning task with a WM task. Participants learned by reward feedback to map a stimulus with two feature dimensions onto one of two response keys. They were also required to hold the stimulus in working memory, for delayed recall. Crucially, either individual features (shape or colour) or bound objects (the combination of shape and colour) were relevant for the learned mapping. Concurrently, the memory task was framed either in terms of single features or bound objects. The optimal strategy to learn feature-based rules uses feature-based representations, whereas object-based rules are best learned using object-based representations. We found that framing working memory in terms of objects or features biased reinforcement learning. Feature-based WM framing biased learning towards the use of features, whereas object-based WM framing biased learning towards the use of objects. The format of visual WM, therefore, governed how the visual feature space was used for reinforcing actions. Our results demonstrate that WM representations are used for learning and can be flexibly shifted depending on the framing of the memory task. This provides evidence for the central role of WM in providing representations that are harnessed by other cognitive systems such as reinforcement learning.