We introduce a machine-learning approach to predict the complex non-Markovian dynamics of supercooled liquids from static averaged quantities. Compared to techniques based on particle propensity, our method is built upon a theoretical framework that uses as input and output system-averaged quantities, thus being easier to apply in an experimental context where particle resolved information is not available. In this work, we train a deep neural network to predict the self intermediate scattering function of binary mixtures using their static structure factor as input. While its performance is excellent for the temperature range of the training data, the model also retains some transferability in making decent predictions at temperatures lower than the ones it was trained for, or when we use it for similar systems. We also develop an evolutionary strategy that is able to construct a realistic memory function underlying the observed non-Markovian dynamics. This method lets us conclude that the memory function of supercooled liquids can be effectively parameterized as the sum of two stretched exponentials, which physically corresponds to two dominant relaxation modes.