A group of scientists in Germany and China has developed a new type of memory chip that could make artificial intelligence (AI) much smarter—and safer.
These tiny components, called memristors, work in a way that’s similar to brain cells. They could help solve one of AI’s biggest problems: “catastrophic forgetting.”
When current AI systems learn something new, they often forget what they learned before. This is because each new task overwrites the previous one.
But our brains don’t work like that. They can learn new things without forgetting old memories.
That’s because our brain cells can adjust how strongly they change with each new experience—a feature scientists call metaplasticity.
Inspired by this, researchers led by Ilia Valov at the Jülich Research Centre in Germany created a new type of memristor.
These devices can switch between analog and digital modes and work across a wide range of voltages.
Most importantly, they can keep old information while learning new things—just like our brains.
Memristors are special because they “remember” past electrical signals. Unlike regular computer parts, their resistance stays the same even after the power is turned off. This memory happens because of tiny physical changes inside the material—like atoms moving around or forming small bridges between parts of the device.
Until now, memristors had some problems. They were fragile, had a short life, and were easily damaged by heat or pressure.
Valov and his team solved this by creating a new switching method called filament conductivity modification (FCM).
Instead of relying on unstable metal filaments or high-voltage oxygen movement, their memristor uses a stable filament made of metal oxides, especially tantalum and oxygen. This filament never fully disappears—it just changes slightly, making the device much more durable.
This new design is more reliable, lasts longer, and works well at lower voltages. That means it’s easier to make and less likely to fail during production.
Another big advantage: these memristors can work both digitally (like standard computer bits) and in analog mode (handling many in-between values). This flexibility is perfect for neuromorphic chips—computers designed to mimic how our brains work. And because the memristors don’t forget older information when learning new things, they could help AI become more stable and intelligent.
In early tests using computer simulations, the new memristor-based system performed very well at recognizing images.
In the future, the team hopes to find even better materials to keep improving these brain-like chips.
Their work, recently published in Nature Communications, could lead to big breakthroughs in how computers learn and remember.
