Temporal Memory Networks with Hamiltonian Evolution: Toward Human-Like AI
Year: 2025
Type: article
Abstract: © 2025 Nathaniel MacDonald. All rights reserved. This work presents Temporal Memory Networks (TMemNet), a novel AI architecture grounded in Hamiltonian evolution principles for continual learning and dynamic memory retention. Unlike conventional deep learning models that rely on static representations or recurrent architectures prone to forgetting, TMemNet introduces a modified Hamiltonian framework to maintain structured energy conservation across time. Key innovations include: Timefield Evolution Constraint: A Hamiltonian-based memory update rule that preserves past knowledge while integrating new information. Orthogonal Projection for Memory Stability: Prevents catastrophic forgetting through structured learning dynamics. Cross-Domain Generalization: Demonstrated superior transfer learning in CIFAR-100, MNIST, SVHN, and Moving MNIST benchmarks, outperforming traditional models in future frame prediction and continual adaptation. This research bridges physics-based AI modeling and lifelong learning, offering a framework that more closely mimics human memory dynamics than existing state-of-the-art architectures such as Transformers and ConvLSTMs. Any reproduction, distribution, or derivative works must credit the author and obtain prior permission.
Author: MacDonald, Nathaniel
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6967959722
FWCI:
Cites:
Cited by:
Related to: