Memory-Decay-Based AI Reasoning for Human-Like Forgetting Systems

Abstract

Catastrophic forgetting in artificial neural networks is a basic obstacle for continual, lifelong learning systems. This paper draws on existing works, particularly those developed within the past few decades, in the memory-decay-based artificial intelligence (AI) reasoning field inspired by human cognitive memory models, such as Ebbinghaus forgetting curves, working memory decay mechanisms, and hippocampal-neocortical complementary learning systems, up to 2020. In parallel, a comparative analysis of regularisation-based, rehearsal-based, and generative replay methods is carried out, based on quantitative benchmarks on Split-MNIST, Permuted-MNIST, and Split-CIFAR10 datasets. Results show that elastic weight consolidation (EWC) preserves accuracy up to 97.2% for sequential tasks, and architectural methods like REMIND outperform the vanilla stochastic gradient descent (SGD) baselines by more than 70 percentage points on Split-CIFAR10. The synthesis also indicates that when optimised computationally, it is possible to obtain significantly better long-term retention by using a biologically-inspired spaced repetition schedule compared with a massed-practice schedule, by about 45–55%. This paper proposes seven detailed tables and figures that represent the dynamics of forgetting, benchmark performances, and cognitive-computational parallels, providing an integrated theory for human-like forgetting systems in AI.