Long-history recommenders often compress a user’s full sequence into a candidate‑independent memory that is cached and reused to score large candidate pools. Real user sequences exhibit multi‑scale semantic structure, with short‑term intent, medium‑term interests and long‑term preferences coexisting. A single cached memory preserves these scales unevenly; linear probes recover recent and mid‑range content far worse than long‑range content, a failure mode termed temporal aliasing. MARS introduces a multi‑resolution user memory that writes the full history into several recurrent state tracks anchored to different half‑lives. A sparse routing reader materializes compact seed memories by selecting the appropriate temporal resolution for each seed, keeping candidate scoring fixed‑size. Experiments on three public datasets show consistent gains over strong baselines, with improvements growing with history length. Component‑matched ablations reveal that temporal diversity and selective routing each add value beyond hard‑window memories or extra capacity. Compared with an interface‑matched baseline, MARS’s advantage widens after within‑user behavioral shifts, with warm‑cache serving latency for 1,000 candidates about 1.02× that of the baseline.
Overall, MARS mitigates temporal aliasing through multi‑scale memories and sparse routing, offering a scalable solution for long‑sequence recommendation. Review