Researchers at the University of Cambridge published a Science Advances paper (DOI: 10.1126/sciadv.aec2324) describing a modified hafnium oxide memristor that switches at heterointerfaces rather than forming unreliable filaments, yielding switching currents approximately one million times lower than conventional oxide-based memristors and hundreds of stable analog conductance levels. The device supports spike-timing dependent plasticity, making it compatible with neuromorphic learning, and maintained stability across tens of thousands of cycles. The team estimates the approach could reduce AI inference energy consumption by up to 70%, though current fabrication requires ~700°C processing temperatures.