Human-like Episodic Memory for Infinite Context LLMs

• DOI: 10.48550/arxiv.2407.09450
• Publication Date: 2024-07-12
• Authors: Zafeirios Fountas; Martin A Benfeghoul; Adnan Oomerjee; Fenia Christopoulou; Γεράσιμος Λάμπουρας; Haitham Bou-Ammar; Jun Wang
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Computer Science - Computation and Language; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Quantitative Biology - Neurons and Cognition; FOS: Biological sciences; Neurons and Cognition (q-bio.NC); Computation and Language (cs.CL); Machine Learning (cs.LG)

Large language models (LLMs) have shown remarkable capabilities, but still struggle with processing extensive contexts, limiting their ability to maintain coherence and accuracy over long sequences. In contrast, the human brain excels at organising retrieving episodic experiences across vast temporal scales, spanning a lifetime. this work, we introduce EM-LLM, novel approach that integrates key aspects of memory event cognition into LLMs, enabling them effectively handle practically infinite context lengths while maintaining computational efficiency. EM-LLM organises sequences tokens coherent events using combination Bayesian surprise graph-theoretic boundary refinement in an on-line fashion. When needed, these are retrieved through two-stage process, combining similarity-based temporally contiguous retrieval for efficient human-like access relevant information. Experiments on LongBench dataset demonstrate EM-LLM's superior performance, outperforming state-of-the-art InfLLM model overall relative improvement 4.3% various tasks, including 33% PassageRetrieval task. Furthermore, our analysis reveals strong correlations between segmentation human-perceived events, suggesting bridge artificial system its biological counterpart. This work not only advances LLM capabilities extended contexts also provides framework exploring mechanisms, opening new avenues interdisciplinary research AI cognitive science.