MemoRAG is a groundbreaking RAG framework that enhances evidence retrieval by leveraging a memory model to achieve a global understanding of massive databases. It efficiently handles up to 1 million tokens in a single context with features like contextual clues and efficient caching. The project aims to achieve light-weight optimization and enhance memory capabilities for diverse applications. Memory-augmented retrieval functionality is also provided for improved evidence retrieval. If you want to try MemoRAG, a toy demo is available. Recent LLMs are supported as memory models for enhanced context length. Evaluation results show MemoRAG’s superior performance over other models. For usage and evaluation, refer to the detailed instructions provided.
https://github.com/qhjqhj00/MemoRAG