Neurobioreasonablely Inspired Long-Term Memory for Large Language Models

View a PDF of the paper titled HippoRAG: Neurobioreasonablely Inspired Long-Term Memory for Large Language Models, by Bernal Jim’enez Guti’errez and 4 other authors

View PDF
HTML (experimental)

Abstract:In order to thrive in unfriendly and ever-changing authentic environments, mammalian brains carry ond to store huge amounts of understandledge about the world and continupartner fuse recent alertation while dodgeing catastrophic forgetting. Despite the astonishive accomplishments, huge language models (LLMs), even with retrieval-augmented generation (RAG), still struggle to efficiently and effectively fuse a huge amount of recent experiences after pre-training. In this labor, we begin HippoRAG, a novel retrieval summarizelabor encouraged by the hippocampal indexing theory of human lengthy-term memory to assist beginanter and more efficient understandledge integration over recent experiences. HippoRAG synergisticpartner orchestrates LLMs, understandledge graphs, and the Personalized PageRank algorithm to mimic the branch offent roles of neocortex and hippocampus in human memory. We contrast HippoRAG with existing RAG methods on multi-hop ask answering and show that our method outexecutes the state-of-the-art methods retagably, by up to 20%. Single-step retrieval with HippoRAG accomplishs comparable or better executeance than iterative retrieval enjoy IRCoT while being 10-30 times inexpensiveer and 6-13 times speedyer, and integrating HippoRAG into IRCoT transports further substantial gets. Finpartner, we show that our method can tackle recent types of scenarios that are out of accomplish of existing methods. Code and data are useable at this https URL.