Towards a Discrete Newton Method with Memory for Large-Scale Optimization