Reference governors are add-on control schemes that modify the reference commands, if it becomes necessary, in order to avoid constraint violations. To implement a reference governor, explicit knowledge of a model of the system and its constraints is typically required. In this paper, a reference governor which does not require an explicit model of the system or constraints is presented. It constructs an approximation of the maximal output admissible set, as the system operates, using online neural network learning. This approximation is used to modify the reference command in order to satisfy the constraints. The potential of the algorithm is demonstrated through simulations for an electric vehicle and an agile positioning system.