Personalized Retrieval Boosts Long-Term Conversational AI Memory

Long-term conversational AI agents are expected to retain and recall past interactions, but the utility of this memory hinges on retrieving the most relevant information for a specific user at the right time. Existing memory-augmented LLM agents have made strides in creating compact memory banks, yet their retrieval mechanisms often rely on simple query similarity or fixed ranking rules, overlooking the crucial aspect of user-conditioned relevance. To address this limitation, the Profile-guided Personalized Retrieval Optimization (PPRO) framework is proposed. PPRO is a retrieval-centric system designed to make memory retrieval both user-aware and optimizable. It constructs episodic and semantic memory banks from dialogue histories and then generates a user profile based on these accumulated memories. This user profile acts as an explicit personalized prior in memory ranking, allowing the retrieval process to account for stable user attributes, preferences, and relationships. Furthermore, PPRO trains a query rewriter using Group Relative Policy Optimization, leveraging both evidence retrieval quality and the quality of the downstream answer as feedback, all while keeping the core memory banks and answer model fixed. Experiments on LoCoMo and LongMemEval-S datasets demonstrated consistent performance gains over both training-free and training-based baselines, with ablation studies confirming the substantial contributions of both profile-guided ranking and retrieval-oriented rewriting.