Proceedings of the Annual Meeting of the Cognitive Science Society

In machine learning, forgetting is typically treated as a defect, whereas human forgetting is understood as a rational response to limited resources and environmental change. We ask whether forgetting can likewise be optimal for machines under realistic constraints. We formalize a continual learning setting, parameterized by a hazard rate governing task volatility and a memory cost penalizing stored data. We compare three replay-based agents: naïve (no memory), unlimited replay (infinite memory), and resource-rational (finite memory). The resource-rational agent significantly outperforms the unlimited replay agent in overall utility across volatile or memory-constrained environments. Optimal memory usage decreases predictably as task volatility and memory cost increase. The resulting forgetting dynamics are best described by an exponential model, showing quantitative alignment with human memory laws. Forgetting thus provides a rational strategy for machines to use when engaging in continual learning in volatile environments or when subject to memory constraints.