HIERARCHICAL BIO-INSPIRED COGNITIVE MEMORY SYSTEMS: A UNIFIED FRAMEWORK FOR SEQUEN- TIAL INFORMATION PROCESSING AND LONG-TERM BEHAVIORAL PREDICTION

CHEN QINXUE, Fung Wai Kin, Xiangyu Yue

Keywords: Bio-Inspired Learning, Cognitive Architecture, Time Series Learning

Abstract: Human cognition emerges from hierarchical neural architectures that integrate emotion, memory, and temporal processing across multiple timescales—capabilities fundamentally absent in current artificial intelligence systems. Existing approaches, from retrieval-augmented generation frameworks to time-series architectures, operate through static information retrieval or linear temporal processing without the progressive abstraction layers essential for human-like reasoning. Here we introduce a bio-inspired cognitive memory system that transcends these limitations through a unified five-layer hierarchical framework that systematically abstracts information from sensory-level event encoding to meta-cognitive concept formation. Our architecture mirrors the brain's multi-timescale processing organization, implementing selective memory retention through biologically-motivated temporal decay mechanisms while integrating emotion-driven prioritization and circadian modulation. Unlike conventional systems that store unprocessed fragments, our approach employs subject-predicate-object triplets as abstraction carriers, combining enhanced PageRank algorithms with large language models to achieve dynamic memory consolidation that replicates hippocampal-neocortical interaction patterns. We validate this framework across two demanding temporal reasoning domains: financial forecasting using social media sentiment achieves state-of-the-art performance with information coefficients of 0.35 and Sharpe ratios of 5.52, surpassing neural architectures by substantial margins; e-commerce recommendation systems demonstrate perfect hit rates at both Hit@5 and Hit@10 metrics while maintaining NDCG@5 scores of 0.63; mental health screening from conversational data establishes new benchmarks in behavioral pattern recognition and disorder classification. The system's hierarchical abstraction capabilities enable superior long-term prediction across 30-day horizons while maintaining computational efficiency through biologically-inspired compression mechanisms. These results establish a transformative paradigm that bridges neuroscientific principles with practical artificial intelligence applications, offering a scalable framework for human-centered AI systems that maintains consistency with established mechanisms of biological memory processing and neural consolidation.

Primary Area: interpretability and explainable AI

Submission Number: 9368