Recommender systems optimized for short-term engagement often hurt long-term consumer satisfaction and retention. Yet optimizing long-term outcomes is challenging because such signals are sparse and noisy. We study short-form video platforms, where identifying videos that drive retention is particularly difficult because content is short and often passively consumed, making it hard to attribute long-term outcomes to any individual video interaction. We examine whether psychological heuristics can be helpful for long-term optimization in recommender systems. We propose the Attribution-Prediction-Optimization (APO) framework, which uses psychological heuristics to identify the most memorable moments within a session, trains a retention model on this structured signal, and incorporates retention predictions into a multi-objective ranking function. Among several candidate heuristics, the peak–end rule \citep{kahneman1993when} performs best. We implement APO on Facebook Reels. Through a large-scale, long-term field experiment, APO significantly increases Daily Active Users and total sessions on Facebook Reels, with positive spillovers to the broader Facebook App, one of the largest gains observed on the platform and corresponding to approximately \$555 million in annual revenue based on public revenue statistics. Conceptually, our work identifies a boundary condition for \emph{when} behavioral insights improve machine learning (ML): Because ML relies on learning statistical patterns from big data, it naturally favors short-term objectives with high signal-to-noise ratios. In contrast, long-term objectives are sparse and noisy, and benefit from psychological structures to extract meaningful signals beyond what data alone can provide. Therefore, psychological and behavioral theory is particularly valuable for designing personalized AI systems that optimize long-term, rather than short-term, outcomes.