Visual Working Memory (VWM) has been shown to increase encoding efficiency by storing interactive actions as “chunks”. However, it is unclear whether social meanings (e.g., valence) of interactive actions affect the efficiency of VWM encoding. In two experiments, participants studied a memory array with actions that were either involved in threatening or non-threatening interactions. All studies were shown in a point-light display with disconnected joint movements. After studying the array, participants were asked to judge whether a single probe action had appeared in the memory array. We found that participants were better at recognizing actions involved in interpersonal interactions over individual actions. Importantly, participants showed greater sensitivity for identifying actions involved in threatening interactions than in non-threatening interactions. When presented with four pairs of interactions, VWM capacity for encoding threatening actions increased by 27% compared to non-threatening actions with high memory load. Our findings support that VWM is tuned to more efficiently encode threatening actions, and prioritizes stimuli with salient social meanings of valence.