The Beijing Academy of Artificial Intelligence (BAAI) and the Gaoling School of Artificial Intelligence at Renmin University of China have jointly released an innovative AI model framework named MemoRAG. This framework, based on long-term memory, aims to advance the development of Retrieval-Augmented Generation (RAG) technology, enabling it to handle more complex tasks beyond simple question-answering.
MemoRAG employs a novel approach, following the process of "memory-based clue generation – clue-guided information retrieval – retrieval-based content generation," to achieve precise information acquisition in complex scenarios. This technology is particularly suitable for knowledge-intensive tasks in fields such as law, medicine, education, and coding, showcasing significant potential.
The core advantage of MemoRAG lies in its global memory capability, capable of handling single-context data of up to a million words, providing robust support for processing large volumes of data. Additionally, MemoRAG boasts high optimization and flexibility, allowing it to quickly adapt to new tasks and achieve optimal performance. It can also generate precise contextual clues from global memory, enhancing the accuracy of problem-solving and uncovering deeper insights from data.
To support further research and application of MemoRAG, the project team has open-sourced two memory models and provided usage guidelines and experimental results. Experiments show that MemoRAG outperforms baseline models in multiple benchmarks. The BAAI has indicated that while the MemoRAG project is still in its early stages, they look forward to community feedback and will continue to optimize the model's lightweight design, diversity of memory mechanisms, and performance on Chinese corpora.
Technical Report:https://arxiv.org/pdf/2409.05591
Repository:https://github.com/qhjqhj00/MemoRAG