Working memory (WM) theoretically affords the ability to privilege social threats and opportunities over other more mundane information, but few experiments have sought support for this contention. Using a functional logic, we predicted that threatening faces are likely to elicit encoding benefits in WM. Critically, however, threat depends on both the capacities and inclinations of the potential aggressor and the possible responses available to the perceiver. Two experiments demonstrate that participants more efficiently scan memory for angry facial expressions, but only when the faces also bear other cues that are heuristically associated with threat: masculinity in Study 1 and outgroup status in Study 2. Moreover, male participants showed robust speed and accuracy benefits, whereas female participants showed somewhat weaker effects, and only when threat was clearly expressed. Overall results indicate that working memory for faces depends on the accessibility of self-protective goals and on the functional relevance of other social attributes of the face.