“The Missing Memristor Found,” by Dmitri B. Strukov, Gregory S. Snider, Duncan R. Stewart, and R. Stanley Williams of HP Labs, appeared in Nature in May 2008, volume 453, pages 80 to 83. It reported the first physical device behaving as a memristor - short for memory resistor - a circuit element whose resistance depends on the history of current that has flowed through it. In effect, the device remembers how much charge has passed and changes its conductance accordingly.
The memristor had been predicted on theoretical grounds in 1971 by circuit theorist Leon Chua, who argued that alongside the resistor, capacitor, and inductor there ought to be a fourth fundamental two-terminal element linking magnetic flux and charge. For nearly four decades no one built one. The HP team showed that memristance appears naturally at the nanoscale, in their case in a thin film of titanium dioxide sandwiched between platinum electrodes, where the motion of charged atomic vacancies under voltage shifts the film’s resistance and holds it after the voltage is removed.
The reason this matters for AI is that a memristor is, physically, a tunable weight that also stores its own value. A grid of memristors can hold the weight matrix of a neural network and, by Ohm’s law and Kirchhoff’s law, perform a matrix-vector multiplication in a single analog step, with the data sitting exactly where the computation happens. That is the foundation of analog in-memory computing, which sidesteps the constant shuttling of weights between memory and processor that dominates the energy budget of conventional AI chips.
Why business readers should care: the memristor is one of the building blocks behind the search for radically more energy-efficient AI hardware. It is the device-level reason analog in-memory computing is taken seriously, even though manufacturing reliable, precise memristor arrays at scale remains an open challenge.