LoopMemGR: From Behavior Logs to Evolving Memory for Generative Recommendation

2026-07-29

AI Summary

This work addresses a key limitation in existing generative recommender systems, which rely solely on user behavioral history while neglecting the system’s own recommendation decisions and corresponding feedback, thereby failing to reuse preference validation, negative evidence, or exploration signals across requests. To overcome this, we propose LoopMemGR, a closed-loop recommendation experience memory framework that introduces system-side recommendation experience logs to record historical recommendation–feedback trajectories. These logs are distilled into compact experience tokens from three complementary perspectives—recency, frequency, and global context—and integrated into the generative recommendation process. This approach breaks from the conventional paradigm of context modeling based exclusively on user behavior, enabling cross-request accumulation of closed-loop experience and fusion of multi-perspective signals. Experiments on a large-scale industrial Taobao dataset demonstrate that the proposed framework significantly enhances recommendation performance.

Abstract

Generative recommendation formulates next-item prediction as conditional autoregressive generation over discrete Semantic IDs, enabling end-to-end recommendation over large-scale item spaces. However, most existing methods follow a history-as-context paradigm that repeatedly reconstructs user preference from behavior history while discarding system-side recommendation decisions after each request. This creates an asymmetric memory: the system remembers what the user has done, but not what it has previously recommended or learned from the resulting feedback. Consequently, useful preference-validation signals, potential negative evidence, and historical exploration information cannot be directly reused across requests. To address these limitations, we propose LoopMemGR, a closed-loop recommendation experience memory framework for generative recommendation. In addition to the conventional behavior log, LoopMemGR maintains a recommendation experience log that records past recommendation--feedback trajectories. It extracts request-relevant evidence through three complementary views: the recency view captures short-term interaction dynamics, the frequency view summarizes recurring recommendation patterns, and the global view distills transferable regularities shared across users. These signals are compressed into a fixed number of experience tokens to condition the generative backbone under a bounded input budget. Extensive experiments on an industrial Taobao dataset demonstrate the effectiveness of closed-loop experience accumulation and multi-view experience extraction.

Problem

• generative recommendation
• recommendation memory
• behavior logs
• feedback loop
• preference modeling

Innovation

• generative recommendation
• closed-loop memory
• recommendation experience log
• multi-view evidence extraction
• experience tokens