Gradient-Free Learning Resists Catastrophic Forgetting

A new Cognitive Memory Primitive (CMP) architecture learns without backpropagation, showing superior resistance to catastrophic forgetting and challenging conventional AI training paradigms.

Diagram illustrating the Cognitive Memory Primitive (CMP) architecture with its two-tier competitive memory.
Conceptual diagram of the Cognitive Memory Primitive (CMP) architecture.
Visual TL;DR
Catastrophic ForgettingDriver
From the article 2 mentionsCatastrophic forgetting has long been viewed as an intractable training-time defect in neural networks, necessitating complex patching mechanisms like replay or regularization.
Backpropagation Flaw?Context
From the articleThis design directly targets the hypothesis that catastrophic forgetting is a structural consequence of how backpropagation assigns credit, rather than an inherent flaw requiring external fixes.
Rethink LearningContext
challenges conventional AI training paradigms by re-evaluating the fundamental learning process
From the articleThis perspective is challenged by a new architectural approach that re-evaluates the fundamental learning process.
Cognitive Memory PrimitiveCore
new architecture (CMP) learns without backpropagation, departing from traditional gradient methods
From the articleResearchers introduce the Cognitive Memory Primitive (CMP), an architecture that departs from traditional gradient-based methods.
Gradient-Free UpdatesCore
From the articleCMP represents inputs as sparse relational codes, stores them in a two-tier competitive memory, and crucially, learns entirely through local, gradient-free updates.
Sparse Relational CodesCore
From the articleCMP represents inputs as sparse relational codes, stores them in a two-tier competitive memory, and crucially, learns entirely through local, gradient-free updates.
Resists ForgettingOutcome
CMP shows superior resistance to catastrophic forgetting on domain-incremental text protocols
From the article 2 mentionsCatastrophic forgetting has long been viewed as an intractable training-time defect in neural networks, necessitating complex patching mechanisms like replay or regularization.
Outperforms BackpropagationOutcome
empirical validation across 15 text domains shows CMP outperforms backpropagation on forgetting
From the articleThis design directly targets the hypothesis that catastrophic forgetting is a structural consequence of how backpropagation assigns credit, rather than an inherent flaw requiring external fixes.

Catastrophic forgetting has long been viewed as an intractable training-time defect in neural networks, necessitating complex patching mechanisms like replay or regularization. This perspective is challenged by a new architectural approach that re-evaluates the fundamental learning process.

Rethinking Learning: Towards Local, Sparse, Gradient-Free Updates

Researchers introduce the Cognitive Memory Primitive (CMP), an architecture that departs from traditional gradient-based methods. CMP represents inputs as sparse relational codes, stores them in a two-tier competitive memory, and crucially, learns entirely through local, gradient-free updates. This design directly targets the hypothesis that catastrophic forgetting is a structural consequence of how backpropagation assigns credit, rather than an inherent flaw requiring external fixes.

Empirical Validation: CMP Outperforms Backpropagation on Forgetting Resistance

On a controlled domain-incremental protocol across 15 text domains, CMP demonstrated a substantial advantage over a matched-size Transformer trained with online EWC. Specifically, CMP's backward transfer was 15-19x better, a result that remained robust even when the domain order was randomized. This performance gain, however, was accompanied by a real accuracy gap compared to the Transformer baseline. The study also reports a null result on a vision benchmark and an unresolved failure when attempting to integrate CMP with a mechanism for improving raw accuracy, underscoring the nuanced trade-offs and the value of transparent negative results.

© 2026 StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training requires a license. See our terms.
Daniel Singer

Written by

Daniel Singer

Editor, StartupHub.ai

Daniel Singer is the editor of StartupHub.ai, a technology expert and thought leader on AI and its applications across sectors, from fintech and healthcare to developer tooling and consumer software. He writes and tests the tools covered here thoroughly and regularly, and built StartupHub.ai to give founders, operators and buyers a clearer read on what they are actually being sold.