Answers are grounded in this paper and cite their sources.
Bundle recommendation aims to accurately capture user preferences from the complex interactions among users, bundles, and items to generate explainable recommendations. However, existing methods based on large language models (LLMs) treat user interactions as static text, over-looking the dynamic evolution of user preferences. Although recent memory-augmented LLM approaches such as MR.Rec and MemoCRS maintain external memory, their heuristic update rules are non-adaptive and designed solely for in-dividual recommendation, rendering them unsuitable for bundle recommendation scenarios. To address these issues, we propose MERGE, a Memory-Enhanced LLM Agent that autonomously manages memory in response to evolving user behaviors. Specifically, MERGE consists of a Memory Mod-ule and a Reasoning Module. The Memory Module employs a token-based hash structure to dynamically manage up-to-date user profiles through insert, update, forget, summary, and query operations, all of which are triggered by the current user interactions. The Reasoning Module then performs multi-stage reasoning based on retrieved information to generate explainable recommendations. We train MERGE using reinforcement fine-tuning (RFT), which first applies supervised fine-tuning (SFT) to establish a foundational understanding of the task, followed by group relative policy optimization (GRPO) to enhance its ability to invoke the Memory and Reasoning Modules. Experimental results on two benchmark datasets show that MERGE significantly outperforms existing state-of-the-art methods in both recommendation accuracy and explanation quality. We open-source our model and code at https://huggingface.co/no1forever/MERGE.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: