Did you check "differentiable neural computers" by DeepMind [0]? I wonder if that type of model could ultimately automatically infer the things you're describing (since it can do memory reads/writes, execute instructions).