The fourth circuit element that “remembers”

Modern computing is built on a decades-old design where data is stored and processed in different parts of the system. Every time a computer runs an app or loads a file, data has to flow back and forth between the memory unit and the processing unit, creating a bottleneck (called the “von Neumann bottleneck”) that wastes time and energy. This is particularly problematic in the age of AI, which demands extremely fast processing.

One promising solution is to build devices that process and store information in the same place. And that’s where an obscure theoretical idea from 1971 comes in.

Classical circuit theory established four fundamental circuit variables – charge, current, voltage, and magnetic flux (a measure of the magnetic field generated by an electric current). Resistors connect voltage and current, capacitors relate charge and voltage, and inductors link magnetic flux and current. However, the direct connection between charge and magnetic flux remained unaccounted for.

In 1971, Leon O Chua, Professor at the University of California, Berkeley, proposed that there was a missing piece in the circuit. He called it the memristor: a device in which electrical resistance depends not just on the current flowing through it at any moment, but on the history of past electricity that has flowed through it before. In other words, it “remembers” data.

“It should be noted that this paper was published not in a device journal, nor in a science journal, nor in a materials journal. It was in a circuit theory journal,” says Santanu Mahapatra, Professor at the Department of Electronic Systems Engineering (DESE), IISc, emphasising that Leon’s proposal was purely theoretical at that time. “It is like the periodic table. When Mendeleev developed it, he left many things empty, and later, people discovered these elements and filled it. That was Chua’s proposal – that there is something which is not discovered yet.”

Building on this idea of a “missing relationship,” the memristor introduces a fundamentally new behaviour of memory embedded within resistance in circuits. Unlike a conventional resistor, where resistance remains constant, a memristor changes its resistance depending on the history of the charge that has passed through it. Turn the electricity off and on again – and it still remembers what state it was in before. This property is what makes it so potentially powerful for computing.

In 2008, the memristor concept moved from theory to the experimental spotlight after 37 years, when researchers at HP Labs published a Nature paper titled “The missing memristor found”. They claimed to have engineered a nanoscale device exhibiting memristive behaviour. Even if their device didn’t strictly behave like Leon had predicted, the announcement generated enough excitement to position memristors as a potential foundation for next-generation memory and neuromorphic systems (computing systems modelled on the human brain). While there were many criticisms and debates around its development, some of the principles eventually did end up getting used in memristive systems.

Today’s memristive systems are slightly different from the ideal memristor that Leon proposed in 1971 in that they do not necessarily follow his original “missing relationship” mathematical definition. Current flowing through a memristor depends not only on the voltage applied at that moment, but also on the voltages it experienced in the past, giving it a form of memory. This memory effect produces a characteristic hysteresis curve, where the current–voltage plot forms a loop instead of following the same path back and forth, indicating a purely resistive memory dependent on past electrical states.

While scientists continue to debate the theoretical basis of the ideal memristor, its practical versions are already driving advances in next-generation memory technologies. For example, modern memristors underpin technologies such as resistive RAM (RRAM), where metal-oxide-metal structures (such as titanium or tantalum oxides) store information via switching between high- and low-resistance states that remain stable even after the power is turned off.

Material innovation

In conventional electronics, information is stored as charge: the presence of electrical charge represents a ‘1’, its absence a ‘0’. Memristors depart from this paradigm. Instead of storing information as charges, they use physical changes inside the material itself. When a voltage is applied, ions – such as metal atoms or oxygen vacancies (behaving as positively charged defects) – move within the device to respectively form (‘1’) or break (‘0’) tiny conductive filaments, thereby changing its resistance and encoding data.

Early material innovations around 2008 focused on oxide-based systems like titanium oxide, later extending to vanadium and niobium oxides. These devices operate via a reversible “soft breakdown” process where conductive paths are formed and erased without permanently damaging the material. This enables multiple resistance states with many levels between 0 and 1, meaning that a single device could, in principle, store far more information than a conventional memory device.

However, because these mechanisms involve electrochemical processes, they are inherently unpredictable, leading to challenges in reliability and reproducibility.

To address this, newer approaches are exploring alternative material platforms such as ferroelectric, spintronic, multiferroic, 2D, and molecular systems. In materials like 2D semiconductors, for instance, switching can occur through controlled changes in crystal structure or interfaces.

Molecular systems offer another route: one such system has been explored by Sreetosh Goswami, Assistant Professor at the Centre for Nano Science and Engineering (CeNSE), IISc and collaborators, in a paper published in Nature in 2024. In these systems, ions are confined within molecular “cages,” restricting their motion to localised regions rather than long-range migration. This low-entropy environment – creating a more orderly, constrained system – enables highly reproducible switching and allows access to over 16,500 distinct, stable conductance states.

.

What fundamentally determines which materials can reliably store and manipulate information in memristors? Researchers are thinking about this in terms of a material’s thermodynamic landscape – essentially, a map of the energy states that a material can settle into. If a material can support multiple stable states, it can enable multilevel memory. Molecular systems are particularly promising, as their assemblies can exist in many configurations, each representing a distinct state. This insight helps researchers engineer energy barriers and thermodynamic basins to control switching behaviour, noise tolerance, and performance. “We [now] know what we don’t understand,” says Sreetosh, highlighting that while core mechanisms remain unclear, identifying the right questions is now driving more informed material innovation

Another bottleneck with memristors is scalability. To be commercially useful, they must integrate with CMOS (Complementary Metal-Oxide-Semiconductor), the standard platform supporting current memory systems like SRAM, DRAM, and DDR. But because memristors rely on electrochemical processes that are hard to control precisely, fabrication introduces device-to-device variability which makes it difficult to reliably map data onto specific states at scale. “Scalability is a massive challenge,” says Chetan Singh Thakur, Associate Professor at the Department of Electronic Systems Engineering (DESE), IISc. This can limit practical CMOS integration.

While memristors hold significant promise for AI and neuromorphic computing, training models remains a challenge. Modern neural networks are trained using backpropagation – a method that adjusts connections by tracing errors backward through the network. “Nobody has ever realised backpropagation through neuromorphic computing,” says Sreetosh. Nevertheless, memristor-based systems are ideal for inference tasks, where pre-trained models can make fast predictions, particularly in edge devices, such as smartphones, wearables, and sensors.

Beyond computation, memristors’ material-level switching and multilevel conductance states could enable highly sensitive, energy-efficient sensors. “I feel that such materials would be super useful to use for sensing,” says Chetan. This includes sensors used for environmental monitoring, healthcare, and other domains requiring compact, low-power sensing solutions.

Ultimately, the memristor may not replace current memory devices or sensors overnight, but something that began as a simple missing link in traditional circuits has now evolved into a powerful idea that is steadily reshaping how researchers think about memory, device intelligence, and computation itself.