Memory-Aware Multi-Agent Architecture with LLM-based Orchestration for Conversational Recommendation
Summary
- Proposed a memory-aware multi-agent architecture with LLM-based orchestration for conversational recommender systems.
- Integrated long-term memory as a first-class component to influence routing decisions and semantic query formation across sessions.
- Evaluated the framework on the Amazon Reviews dataset, demonstrating recovery of up to 63 percent of the gap toward explicit-query scenarios.
- Showed that supervisory orchestration increases user acceptance by 28.9 percentage points and introduced the Semantic Consistency Score.
Why it matters
System-level design choices like explicit memory integration and LLM orchestration can significantly improve conversational recommendation quality and prevent semantic degradation in extended dialogues.