MEMOIR: Temporal Behavioral Memory for Recommendation Across the Preference-Drift Spectrum

Existing sequential recommendation models often treat user preferences as static or as a byproduct of next-item prediction, failing to explicitly capture how interests evolve over time. MEMOIR (Memory-based Evolution MOdel for Inference in Recommendation) addresses this by treating preference evolution as a first-class architectural component. The framework segments interaction histories into monthly temporal windows, uses an LLM to generate semantic memory embeddings for each period, and employs a novel evolution-preserving contrastive loss to learn the trajectory of these preferences. Finally, it aggregates the user's current state, evolution direction, and predicted future into a single representation for recommendation.

MEMOIR achieves an approximately 18% relative gain in NDCG@10 over the ID-based SASRec baseline. When compared to UniSRec, the strongest baseline, MEMOIR is statistically tied on aggregate ranking metrics. However, a drift-stratified analysis reveals that MEMOIR outperforms UniSRec on ranking quality (NDCG@10, MRR) for users at the high- and low-drift extremes of the population. Conversely, UniSRec performs better on volume-oriented metrics (HR@10, HR@20) and for users in the middle of the drift spectrum. An ablation study indicates that no single component—such as the contrastive loss or temporal segmentation—is solely responsible for the performance gains, suggesting the model's success is tied to the holistic integration of these features.

This paper demonstrates that incorporating explicit temporal evolution into user modeling provides a distinct advantage for specific user segments, particularly those experiencing significant preference shifts. It also highlights a critical challenge in modern recommendation: simply augmenting models with LLM-based semantic features does not automatically improve performance. Without an architectural mechanism to align semantic representations with behavioral co-occurrence patterns, these features can be ineffective or even detrimental. MEMOIR provides a blueprint for how to bridge this gap through trajectory-aware aggregation.