In search of lost memories: Modeling exploration with forgetful generalization

The clarity of our memories guides how we act on past experiences and approach new ones. While extensive research has examined how memory limitations affect recall, far less is known about how these limitations influence decisions in new situations, which depend on generalizing from incomplete or distorted memories. We introduce a computational model of "forgetful generalization", integrating similarity-based generalization with variable-precision memory modulated by recency and asymmetric surprise. In a preregistered spatially correlated bandit experiment, we manipulated (within-subject) whether past observations remained visible (low load) or disappeared (high load). Greater memory load increased "forgetfulness", as captured by parameters indexing recency- and surprise-dependent decay of memory precision. These parameters also predicted individual working memory capacity and explained differences in performance and search patterns. By formalizing how generalization operates under limited memory, our model links episodic reinforcement learning to theories of adaptive memory compression and resource rationality.