We present a data-driven method for the macroscale constitutive modeling of anisotropic microscale progressive damage in composites. Calibrated with path-dependent RVE simulations of damage in a composite microstructure, the constitutive model is formulated as generally as possible while possessing two key characteristics. First, the core of the model consists of one or more small feedforward neural networks, which are calibrated via backpropagation. Second, these neural networks are composed within a stress relation with internal variables such that mechanical dissipation is always non-negative, the model is rate-independent, and proportional strain unloading never results in a change in the damage state. Since the early 2020's, there has been a great deal of interest in the development of data-driven constitutive modeling techniques. The particular challenge presented by path-dependent materials has seen the development of various model regularization techniques to penalize the prediction of stresses resulting in a negative dissipation rate. Much effort has focused on path-dependence resulting from plasticity, but recently some authors have begun developing methods specific to anisotropic progressive damage. We take an alternative approach to these authors to enable hard constraints on the mechanical dissipation and rate independence for arbitrary non-proportional cyclic loading. We will introduce the approach for predicting material point behavior at the macroscale and demonstrate that the model is capable of predicting proportional cyclic loading with limited data and generalizing to complex cyclic loading histories. We will show how the model is embedded within a macroscale finite element analysis and optimization framework to enable microstructure-informed progressive damage analysis and structural optimization.