Reward is not Necessary: How to Create a Compositional Self-Preserving Agent for Life-Long Learning — AI Alignment Forum