MGRetrieval: Memory-Guided Retrieval for Long-Term Dialogue
arXiv paper 2605.27437 introduces MGRetrieval, a retrieval strategy for long-term dialogue agents using Large Language Models (LLMs). It addresses the problem of redundant memory contexts by grounding reflective retrieval in the semantic structure of historical memories. The method consists of two steps: constructing a precise retrieval path by referencing historical memory structure, then performing retrieval. This improves over one-shot retrieval and recent reflection-based methods by reducing instability and latency.
Key facts
- arXiv paper 2605.27437
- Title: MGRetrieval: Memory-Guided Reflective Retrieval for Long-Term Dialogue Agents
- Proposes a retrieval strategy for LLM-based dialogue agents
- Addresses redundant memory contexts in long-term dialogue
- Uses semantic structure of historical memories for retrieval
- Two-step process: construct retrieval path then retrieve
- Improves over one-shot retrieval and reflection-based methods
- Reduces instability and latency
Entities
Institutions
- arXiv