# MemoRAG: Moving towards Next-Gen RAG Via Memory-Inspired Knowledge Discovery

**Authors:** Hongjin Qian, Peitian Zhang, Zheng Liu, Kelong Mao, Zhicheng Dou  
**arXiv:** [2409.05591](https://arxiv.org/abs/2409.05591)  
**Date:** 2024-09-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2409.05591v2/

## Key points

- Authors introduce MemoRAG as a novel retrieval-augmented generation paradigm
- MemoRAG leverages long-term memory to enhance capabilities of large language models (LLMs)
- MemoRAG is a dual-system architecture utilizing light and expensive LLMs for draft and final answers respectively
- Enhances cluing mechanism and increases memorization capacity compared to traditional RAG systems
- Demonstrates superior performance in various tasks through experimental evaluation
- Additional technical details and resources available on authors' GitHub repository: https://github.com/qhjqhj00/MemoRAG

## Summary

, , , , 

In their paper titled "MemoRAG: Moving towards Next-Gen RAG Via Memory-Inspired Knowledge Discovery," authors Hongjin Qian, Peitian Zhang, Zheng Liu, Kelong Mao, and Zhicheng Dou introduce MemoRAG as a novel retrieval-augmented generation paradigm that leverages long-term memory to enhance the capabilities of large language models (LLMs). MemoRAG is a dual-system architecture that utilizes both light and expensive LLMs to generate draft and final answers respectively. It also improves upon traditional RAG systems by enhancing its cluing mechanism and increasing its memorization capacity. Through experimental evaluation, MemoRAG demonstrates superior performance in various tasks compared to conventional RAG systems. The authors provide additional technical details and resources for accessing codes and models related to MemoRAG on their GitHub repository (https://github.com/qhjqhj00/MemoRAG). This innovative approach showcases the potential of integrating long-term memory into retrieval-augmented generation for advancements in natural language processing and knowledge discovery.