Hypergraph Memory for Multi-step RAG

This paper addresses the limitations of existing memory mechanisms in multi-step retrieval-augmented generation (RAG) systems. It proposes a hypergraph-based memory (HGMem) to capture high-order correlations between facts, leading to improved reasoning and global understanding in long-context tasks. The core idea is to move beyond passive storage to a dynamic structure that facilitates complex reasoning and knowledge evolution.

• •Proposes HGMem, a hypergraph-based memory mechanism for multi-step RAG.
• •HGMem captures high-order correlations between facts.
• •Improves reasoning and global understanding in long-context tasks.
• •Outperforms strong baseline systems on challenging datasets.