Reproduce “Semantic XPath: Structured Agentic Memory Access for Conversational AI” on OpenResearch

About this paper

Yifan Simon Liu, Ruifan Wu, Liam Gallagher, Jiazhou Liang, Armin Toroghi, Scott Sanner

Researchers at the University of Toronto developed SEMANTIC XPATH, a tree-structured memory module for Conversational AI, which uses an XPath-style query language with semantic capabilities to access and update structured conversational memory. This system achieves pass rates comparable to in-context memory while consuming substantially fewer tokens, and significantly outperforms flat RAG methods, particularly in multi-turn interactions.

Key insights

• Integrating semantic understanding with XPath-style structured querying allows for precise and contextually appropriate retrieval from hierarchical memory, outperforming flat RAG methods.
• Retrieving only relevant substructures from a tree-structured memory module enables stable and reduced token consumption across long multi-turn conversations, addressing scalability issues of in-context memory.
• The use of an entailment-based semantic scorer for query interpretation generally leads to higher retrieval accuracy compared to a cosine similarity-based scorer.

Summary by alphaXiv.