We propose a new framework for univariate Estimation-of-Distribution Algorithms (EDAs) in which neural networks are used at two levels: (i) for generating new solutions from a memory that encodes some information acquired since the beginning of the search, (ii) for updating this memory with the new solutions sampled at each iteration. These neural networks are then learned by neuro-evolution, so as to automatically discover new efficient strategies for solving pseudo-Boolean problems, without having any specific a priori on how to design them. The algorithms discovered demonstrate their competitiveness compared with existing univariate EDAs of the literature. On average, they achieve better scores for the same number of calls to the objective function, when tested on a variety of different distributions.