Flexible goal-directed behavior requires a generalized internal metric that tracks progress toward goals across locations, routes, and behavioral contexts. Although hippocampal–entorhinal cognitive maps encode detailed spatial representations, whether the brain constructs an abstract distance-to-goal state that generalizes across arbitrary start–goal combinations—a prerequisite for reinforcement learning over space—remains unknown. Here we show that the nucleus accumbens (NAc) computes such generalized spatial states. Using large-scale population recordings and causal perturbations in rats navigating toward dynamically changing goals, we find that NAc activity encodes a scale-invariant distance-to-goal signal, normalized by total journey length, across maze geometries and task rules. A decoder trained on this signal transfers to novel journeys, tracking physical path length rather than remaining time. This signal emerges selectively during memory-guided but not cue-guided navigation, dissociating spatial distance-to-goal coding from purely motivational or value representations. It persists during hippocampal and medial entorhinal cortex silencing but depends critically on dopaminergic input from the ventral tegmental area, which is necessary for accurate goal targeting. Beyond the current goal, NAc populations encode distances to previous goals within orthogonal subspaces, enabling parallel evaluation of counterfactual spatial states without interference. Consistent with this multiplexed architecture, reward omission reinstates search behavior toward former goals, whereas suppressing dopamine release in the NAc abolishes this memory-guided reinstatement. These findings establish the NAc as a substrate for generalized, multiplexed spatial state abstraction, revealing a striatal computation that bridges cognitive mapping and reinforcement learning to enable flexible goal-directed behavior in dynamic environments.
