Toward Interpretable and Persistent Personalization: A Memory-Augmented Agent Framework for LLM-Based Travel Planning

With the widespread application of large language models (LLMs) in intelligent conversation and recommender systems, integrating them into travel-related tasks has become a key research focus within the smart mobility domain. However, limitations such as high fine-tuning costs, cold-start challenges, issues in validation and logical coherence, and difficulties in maintaining contextual memory hinder the effectiveness of personalized interactions by traditional LLMs in travel scenarios. To address these challenges, we propose the Reasoning-enhanced Multi-turn Agent with Personalized Adaptation Framework (ReMAP), a generation-augmented agent framework that reduces personalization costs, improves validation and logical interpretability via Reasoning-and-Acting (ReAct) and Chain-of-Thought (CoT) prompting, and incorporates self-updating and retrieval mechanisms for factual memory to enhance the robustness of personalized generation in LLMs. The Tibet tourism-oriented personalized interaction agent system built upon this framework demonstrates strong performance in multiple-round, multi-group response experiments conducted under real-world travel scenarios. Experimental results show that ReMAP significantly outperforms baseline approaches in cold-start responsiveness (+10.51% personalization accuracy), itinerary feasibility (+12.38% pass rate), and long-term personalization consistency.