EMM: Plug-and-play memory-bank sampling for contrastive recommendation

In contrastive recommendation, representation quality depends not only on the encoder but also on sample construction: positives drive alignment and negatives enforce uniformity, shaping gradient strength, optimization stability, and data efficiency under sparse feedback. However, two key challenges have not been well explored in existing solutions: (i) how to obtain negatives that are simultaneously informative, fresh, and diverse, since random sampling yields weak gradients while static hardest only mining quickly becomes stale; (ii) how to safeguard against false negatives, where semantically close items are mistakenly penalized, distorting geometry and destabilizing training. To cope with these challenges, we propose Episodic Memory Mining (EMM), which reframes negative sampling from a brittle heuristic into a principled component for contrastive recommendation, remaining plug-and-play by intervening only at the sampling layer and keeping backbones intact. EMM keeps negatives fresh, informative, and diverse via an EMA memory with periodic rebuild, Top-K retrieval from a refreshed candidate pool, and a small random-negative mix with rolling refresh to stabilize gradients. Across three benchmarks and five backbones, EMM yields about 15.1% average improvement in overall recommendation quality at comparable training cost and shows broader hyperparameter robustness. This supports negative sample construction as a simple and general lever for stronger contrastive recommendation under sparse feedback.