A Dual-Process Memory-Augmented Large Language Model Agent for Explainable Sequential Recommendation

Abstract

Large language model (LLM) recommenders have progressed from treating recommendation as a language task to full-fledged recommendation agents organized around profile, memory, planning, and action modules. Two frontiers have recently emerged: making an agent’s memory self-evolving, and injecting reasoning through reinforcement fine-tuning. Yet two problems remain poorly addressed. First, an efficiency–coverage tension: scoring or generating over a million-item catalogue with an LLM is prohibitively expensive and invites hallucinated (non-existent or irrelevant) items. Second, a memory–reasoning disjunction: most systems either accumulate memory or perform single-shot reasoning, lacking a principled controller that decides when to recall quickly and when to deliberate slowly, so accuracy and explainability are hard to obtain jointly. Inspired by the dual-process theory of cognition, we propose DUET, a fast–slow cooperative recommendation agent with three components. 1) A System-1 fast recall / System-2 slow reasoning schedule: a lightweight two-channel retriever pre-selects a small candidate set from the full corpus, and the LLM performs chain-of-thought re-ranking and explanation only over that set, sidestepping full-catalogue generation. 2) A hierarchical self-evolving memory that separates short-term session intent from long-term preference and uses a reflect-and-write-back mechanism to consolidate memory, mitigating drift on long histories. 3) A confidence-gated deliberation rule that invokes System-2 reasoning and self-reflection only when retrieval confidence is low or candidates are homogeneous, realizing on-demand thinking. Beyond Recall@ $K$ and NDCG@ $K$ , we report explanation faithfulness and average tokens per recommendation. On Amazon Reviews 2023, MovieLens-1M and MIND, DUET improves Recall@10 by 6.1% to 7.1% and NDCG@10 by 7.6% to 8.2% over the strongest baseline while using a small fraction of the LLM tokens consumed by prior agentic recommenders, and yields the most faithful explanations. A human annotation study over $1{,}200$ explanations corroborates the automatic faithfulness measures, and a complete specification of prompts, decoding, serving and training configurations is provided so that every reported number can be reproduced.