Bits manipulation in traditional memory writing is commonly done through quasi-static operations. While simple to model, this method is known to reduce memory capacity. We demonstrate how a reinforcement learning agent can exploit the dynamical response of a simple multi-bit mechanical system to restore its memory. To do so, we introduce a model framework consisting of a chain of bi-stable springs manipulated on one end by the external action of the agent. We show that the agent learns how to reach all available states for three springs, even though some states are not reachable through adiabatic manipulation, and that training is significantly improved using transfer learning techniques. Interestingly, the agent also points to an optimal system design by taking advantage of the underlying physics. Indeed, the control time exhibits a non-monotonic dependence on the internal dissipation, reaching a minimum at a cross-over shown to verify a mechanically motivated scaling relation.