To successfully navigate their environment, animals may generate an internal, abstract and stable representation of the environment that can be updated based on sensory cues, such as visual landmarks or vestibular feedback, or internally generated motor commands. In mammals, neurons that fire when the animal faces a particular direction in the world reference frame have been recorded in mammillary bodies and entorhinal cortex. The heading direction can be read out using a weighted sum of the activation of those cells. This dynamics is well described by a “ring attractor” model: the network has an attractor with a ring topology in its phase space, positions along the ring represent the heading direction, and input signals translate the network activation along this circular attractor. Although this model has found remarkable validation in the Drosophila ellipsoid body, the precise anatomical dissection of a ring attractor circuit has been elusive in the vertebrate brain. Here we describe a group of ~100 GABAergic neurons in the anterior hindbrain (aHB) of the larval zebrafish with highly constrained dynamics lying on a ring manifold in phase space. Clockwise and counterclockwise shifts along this manifold happen when the fish performs left and right movements, so that the network phase keeps track of current heading direction. This heading representation is stable over tens of minutes even without external visual or vestibular feedback. We show that neurons are anatomically organized according to their proximity in activity space and connect with each other in the interpeduncular nucleus (IPN), a structure implicated in navigation in fish and mammals. Moreover, their morphologies suggest a mechanistic model for the organization of the ring network dynamics. Together, our data represent the first observation of a full ring attractor network with an anatomical organization that integrates heading direction in the vertebrate brain.