Neuromorphic devices Novel memristor aims to prevent "catastrophic forgetting" in AI

Related Vendors

Forschungszentrum Jülich GmbH

Wuxi InfiMotion Technology Co., Ltd.

A team from the Jülich Research Center has developed a new type of memristor designed to prevent energy-efficient AI chips from "catastrophically forgetting" when transitioning from one AI model to the next.

Hardware for artificial neural networks has a fundamental problem: agentic AI uses optimized models for various tasks, but when switching from one model to another, there is a risk that all data is lost—already learned or generated content is lost in the memory of AI agents. This phenomenon is also referred to in AI research as "catastrophic forgetting."

Memristors like ReRAM can drastically reduce the power consumption of edge AI chips through in-memory processing, but still face challenges when switching between models. Reliable arrays of these memristors can store both the inference weights for AI models and the hidden weights used across models.

A team from the Jülich Research Center has now introduced a new type of memristor that could counteract this effect. The component can be selectively modulated without completely overwriting stored states— similar to the human brain with different degrees of synaptic change.

New switching mechanism with high stability

"We have discovered a fundamentally new electrochemical memristor mechanism that is chemically and electrically more stable," explains Prof. Ilia Valov, head of the research group. "Its unique properties allow the use of various switching modes to control the modulation of the memristor so that the stored information is not lost," he says.

Technically, the memristor is based on a novel mechanism: instead of relying solely on metallic filaments (ECM) or oxygen ion-based valence changes (VCM) as before, it is based on a mixed oxide filament formed from oxygen and tantalum ions. This filament remains permanently and is merely chemically modified—an approach that the researchers refer to as "Filament Conductivity Modification" (FCM).

The result is a robust, chemically and electrically stable memristor that operates at lower voltages and is less susceptible to thermal stress or mechanical influences. This reduces the rejection rate in manufacturing, and the components have a longer lifespan—important criteria for industrial use.

Memristors for AI

Can AI chips get a sense of time?

Analog meets digital—ideal for neuromorphic systems

Another advantage lies in the flexible operation: the new memristor can be operated both analog and digital. While digital states provide distinct memory values, analog signals allow for finely graded weightings—central for mapping learning processes. This is exactly what makes the component particularly interesting for "computation-in-memory" architectures, where computing and storage functions are integrated.

In simulations with artificial neural networks, the memristor already showed promising results in pattern recognition in image data. The technology could play a key role in the development of energy-efficient, learning-capable hardware in the future—such as in autonomous systems, edge AI, or embedded sensors.

The researchers now want to explore more materials to further improve stability and switching properties. The goal is scalable integration into existing CMOS manufacturing processes. "Basic research is essential to better control processes at the nanoscale," says Valov, who has been working in the field of memristors for many years. "We need new materials and switching mechanisms to reduce the complexity of the systems and increase the range of functionalities."