Reproduce “AtomRec: Evolving Atomic Memory for Agentic Recommendation” on OpenResearch

About this paper

Peiyu Hu, Weihai Lu, Siying Gu, Zhuodong Liu, Zhaokai Luo, Yuean Niu, Zhiyong Wang, Jia Wang

AtomRec stores specific preference traces as structured notes, links notes through semantic relationships, evolves relevant fields as new evidence arrives, and retrieves linked paths for recommendation. In an illustrative Amazon Books case, it ranks On the Edge of the Etheric first, unlike reduced-memory versions. Across four benchmarks, it reports around 8.5% average relative improvement, including N@5 of 0.7130 versus 0.6480 for MemRec; this is ranking evidence under sampled offline evaluation, not proof of satisfaction or latency.

Key insights

• Field-level memory evolution preserves temporal evidence while allowing earlier preference descriptions to be refined as new interactions arrive.
• Semantic links contribute more than embedding similarity alone by expressing relationships such as shared themes, intent progression, preference transitions, and cross-domain transfer.
• The approach is particularly helpful under preference drift, with its relative advantage over MemRec increasing from 5.2% for low-drift users to 13.9% for high-drift users.

Summary by alphaXiv.