Synaptic plasticity allows animals to adapt to the environment. However, making permanent synaptic changes requires a significant amount of metabolic energy. This cost is so high that learning reduces the lifespan of fruit flies by 20% when feeding is stopped (Mery and Kawecki, 2005). Thus the brain should carefully regulate learning. For instance, flies stop some forms of memory formation to survive upon starvation (Placais and Preat, 2013). To examine when it is best to halt energy-costly learning, we used a computational reinforcement learning model which takes the animal’s energy budget into account. In the model, flies should learn to avoid the hazard from aversive stimuli. However, this consumes energy and exposes them to starvation hazard. We implemented a high-cost long-term memory (LTM) pathway and a low-cost, but less persistent, anesthesia-resistant memory (ARM) pathway, and find an energy efficient learning policy by exploring how the brain switches memory pathways to maximize survival. Consistent with experimental results (Placais and Preat, 2013), the lifespan in our model is prolonged when LTM is gated by energy reserve. Moreover, we find that it is more energy efficient to learn by depressing the weight inducing the unwanted action than by potentiating the weight of the desired action, again consistent with experiments (Perisse et al., 2016). We propose that energy considerations pervade learning and memory across species.