A memristor is an electronic component that changes its resistance depending on how much electric charge has flowed through it, and it remembers that resistance even after the power is turned off. The name is short for “memory resistor.” First predicted on paper in 1971 and not built in a lab until 2008, the memristor is often called the fourth fundamental circuit element, joining the resistor, capacitor, and inductor. It has attracted enormous interest because it could reshape how computers store data and process information.
Why It Was Predicted Before It Existed
In 1971, electrical engineer Leon Chua at UC Berkeley noticed a gap in circuit theory. Three fundamental components already existed, each defining a relationship between pairs of basic electrical quantities (voltage, current, charge, and magnetic flux). But one relationship had no matching component. Chua used symmetry arguments to propose that a fourth element should exist, one linking charge to magnetic flux. He called it the memristor and showed mathematically that it would have valuable circuit properties, including the ability to “remember” its electrical state.
For 37 years, the memristor remained purely theoretical. No one could point to a physical device that behaved the way Chua described. Then in 2008, a team at HP Labs announced they had built one using a thin film of titanium dioxide sandwiched between two electrodes. That demonstration confirmed Chua’s prediction and kicked off a wave of research that continues today.
How a Memristor Works
A typical memristor is a thin sandwich: two metal electrodes with a layer of metal oxide between them. When you apply a voltage, tiny charged defects called oxygen vacancies migrate through the oxide layer. These vacancies form conductive pathways, or filaments, between the electrodes. The more filaments that form, the lower the device’s resistance. When the voltage reverses, the filaments partially dissolve and resistance climbs back up.
The key feature is that when you remove the voltage entirely, the filaments stay wherever they were. The device holds its resistance state without any power, which is what makes it a form of non-volatile memory. Research using advanced X-ray analysis has confirmed that even after switching the device off, some oxygen vacancies remain in position, showing genuine retention of the previous state.
If you plot the current flowing through a memristor against the voltage applied to it, you get a distinctive figure-eight shape called a pinched hysteresis loop. The loop crosses the origin (zero voltage, zero current) but traces a different path on the way up than on the way down. That enclosed area is a direct signature of the memory effect. At very high frequencies the loop collapses and the device acts like an ordinary resistor, because the oxygen vacancies can’t migrate fast enough to respond. At lower frequencies, vacancies have more time to redistribute, and the memory effect becomes more pronounced.
Speed and Energy Consumption
Commercially available memristors today switch states in roughly 50 nanoseconds to 100 microseconds, which is slower than state-of-the-art DRAM. But laboratory devices have pushed far beyond that. Researchers have built ultra-fast memristors from atomically thin sheets of hexagonal boron nitride (a two-dimensional material) that switch in just 120 picoseconds, less than a billionth of a second, with a switching energy of only 2 picojoules per operation. That speed falls squarely within the range modern processor circuits already operate in, making these devices potentially compatible with existing chip designs.
Low energy consumption is one of the memristor’s most appealing traits. Because switching involves moving a small number of ions rather than charging large capacitors, each operation can require very little power. For battery-powered sensors, wearable devices, and spacecraft electronics, that efficiency matters enormously.
Why Memristors Matter for AI Hardware
Conventional computers shuttle data back and forth between a processor and separate memory chips. This constant data movement creates a speed bottleneck and wastes energy. Memristors offer a fundamentally different approach: they can store data and perform calculations in the same physical location. Arranging memristors in a grid (called a crossbar array) lets the array multiply matrices, the core mathematical operation behind neural networks, in a single step rather than thousands of sequential instructions.
Even more intriguing, memristors behave remarkably like biological synapses. A synapse strengthens or weakens depending on how often it’s stimulated. Memristors do the same thing: repeated electrical pulses gradually shift their resistance, mimicking both short-term plasticity (temporary changes) and long-term plasticity (lasting changes). Researchers have demonstrated memristors that reproduce several key synaptic behaviors, including paired-pulse facilitation (where two pulses in quick succession produce a stronger response than either pulse alone) and spike-rate-dependent plasticity (where the frequency of stimulation determines how much the connection changes). In neural network simulations, one such device achieved 94.6% pattern recognition accuracy.
This synaptic resemblance makes memristors natural building blocks for neuromorphic computing, hardware designed to mimic the brain’s architecture rather than follow the traditional processor-plus-memory model.
Memristors as Next-Generation Memory
Memristor-based memory goes by several names in the industry, most commonly resistive RAM or RRAM. Compared to traditional flash memory, RRAM offers several structural advantages: a simpler device design, compatibility with standard chip manufacturing processes, the ability to stack in three dimensions for extremely high density, fast read speeds, and low operating voltage. RRAM also shows strong endurance through many write-erase cycles and excellent radiation hardness, a property that makes it attractive for satellites and space missions where conventional flash memory can be corrupted by cosmic rays.
That said, RRAM has not yet displaced established technologies like DRAM, SRAM, or flash at scale. Cost and consistency remain challenges. The technology currently fills specialized niches rather than serving as a drop-in replacement for mainstream memory.
Flexible and Organic Memristors
Not all memristors are rigid silicon-based devices. Researchers have built organic memristors on flexible plastic substrates using solution-based manufacturing, essentially printing the devices onto bendable films. These flexible memristors retain their switching behavior even when bent, making them candidates for wearable electronics and smart sensors that need to conform to the body. One research group demonstrated a flexible artificial synapse capable of solving complex optimization problems, pointing toward wearable neural networks that could run AI tasks directly on the device rather than relying on a cloud connection.
Who Makes Them Today
Several companies have moved memristor technology from the lab toward commercial availability. Knowm Inc. began selling self-directed channel memristors in 2015, making it one of the first to offer memristor chips for prototyping and research, including crossbar arrays designed for neuromorphic applications. Weebit Nano focuses on RRAM that is compatible with standard chip fabrication, targeting the Internet of Things and embedded systems. Fujitsu sells RRAM chips under the brand name RAMXEED, aimed at low-power applications in connected devices. HP Labs, despite pioneering the first physical memristor in 2008, fell behind in commercialization and no longer leads the market.
The technology remains early-stage in commercial terms. Most available products serve researchers, prototype developers, and specialized industrial applications rather than consumer electronics. But the path from theoretical curiosity to purchasable chip has already been traveled, and the range of applications, from AI accelerators to radiation-hardened space memory to flexible wearable devices, continues to expand.