Navigating towards goals in crowded spaces is a complex behavior that requires keeping track of relative positions of goals and obstacles. Animals can solve this by breaking down routes into subgoals - intermediate waypoints. It is not known how the brain represents subgoals, or how their representation differs from static spatial features of the environment. Building on work showing that mice use a subgoal strategy to escape to shelter, we placed mice in an arena divided in half by a barrier, with a shelter in the bottom half. The barrier leaves a gap on one side which mice learn to use as a subgoal when escaping to the shelter from a threat in the top half. This design decouples subgoals from static spatial features, as the open edge is only behaviorally relevant when the mouse is in the top half. We then flip the gap to the opposite side, forcing a subgoal update and allowing us to track how the spatial and value representations develop. We recorded single unit activity in the Retrosplenial Cortex (RSC), an area that encodes the head-angle to shelter and is necessary for shelter-directed escapes. Population analyses show high decoding accuracy for the open barrier edge in a dynamic manner that follows the barrier flip and is correlated with behavior. After the flip, the need to learn the new location and unlearn the previous one slows the representation update. This representation of the subgoal is built from single neurons dynamically tuned to the open edge, dependent on the arena configuration and the mouse location relative to the subgoal. These data show that RSC neurons encode subgoals, revealing a spatial representation dynamically updated over two timescales: contextual value learned over minutes, and positional value learned over seconds. This representation type might extend to other compositional goal-directed tasks.