In the era of rapidly expanding digital economies and information technologies, data has become a critical resource, driving unprecedented demands for high-capacity, low-latency, and energy-efficient storage solutions. Traditional von Neumann computing architectures, characterized by the physical separation of memory and processing units, face inherent bottlenecks in data movement and power consumption, particularly under the exponential growth of artificial intelligence (AI), the internet of things (IoT), cloud computing, and autonomous systems. Addressing these challenges requires novel non-volatile memory technologies that combine high density, fast switching, low power consumption, and compatibility with in-memory computing. Among these, resistive random-access memory (RRAM) has emerged as a leading candidate due to its simple device structure, high scalability, multi-level storage potential, and CMOS process compatibility.

This review systematically summarized recent progress in RRAM research, focusing on physical mechanisms, performance optimization, array integration, and application-driven development. At the device level, RRAM operates through reversible resistive switching between high-resistance (HRS) and low-resistance (LRS) states. Its dominant mechanisms include electrochemical metallization (ECM), valence change (VCM), and thermochemical effects, which govern the formation and rupture of nanoscale conductive filaments. ECM devices rely on metal ion migration and redox reactions, offering large switching windows but lower uniformity, whereas VCM devices, driven by oxygen vacancy dynamics, exhibit superior endurance and reproducibility but with narrower switching ranges. Quantum-scale effects, including filament quantization and electron-ion coupling, become increasingly significant as device dimensions shrink to atomic levels, influencing stochasticity, variability, and long-term reliability. Advanced theoretical and computational tools—first-principles calculations, multiscale simulations, and compact device modeling—have enabled deeper understanding of these mechanisms, providing a foundation for rational device design and materials engineering.

Performance modulation strategies are central to realizing RRAM's practical potential. Material doping, interface engineering, and electrical programming collectively enhance switching uniformity, multi-level storage capability, and endurance. Transition metal and nitrogen doping, alloying, and hybrid organic-inorganic compositions optimize oxygen vacancy formation and conductive filament evolution, while multilayer interfaces and buffer layers regulate local electric fields and defect distributions, suppressing random filament nucleation and stabilizing switching dynamics. Pulsed electrical programming, involving careful control of amplitude, duration, and waveform, directly modulates ion migration and filament dynamics, enabling precise multi-level conductance states critical for in-memory and neuromorphic computing applications. These strategies highlight the necessity of system-level optimization that balances speed, energy efficiency, and reliability.

At the integration level, array architectures determine the scalability and applicability of RRAM technologies. Crossbar arrays (1R), selector-integrated structures (1S1R), and transistor-assisted cells (1T1R) provide trade-offs between density, uniformity, and reliability, addressing issues such as sneak path currents and read/write disturbances. Three-dimensional (3D) stacking further increases storage density and parallel computation throughput but introduces new challenges in thermal management, interconnect interference, and process complexity. CMOS back-end-of-line (BEOL) integration has enabled embedding RRAM within advanced nodes (e.g., 28 nm, 14 nm), supporting high-density non-volatile memory without modifying front-end logic, while continuing work focuses on material selection, deposition uniformity, and low-temperature process compatibility to preserve device performance and reliability.

RRAM's unique combination of non-volatility, parallel tunability, and integration compatibility has enabled a spectrum of emerging applications. In in-memory computing, RRAM crossbar arrays facilitate direct matrix-vector multiplication, mitigating the von Neumann bottleneck and achieving high energy efficiency in AI acceleration. In neuromorphic computing, RRAM devices emulate synaptic plasticity, supporting spiking neural networks and integrated neuron-synapse architectures for highly efficient pattern recognition and temporal information processing. In intelligent sensing, optoelectronic and chemical RRAM systems integrate sensing, storage, and pre-processing functions, offering high sensitivity, low latency, and energy-efficient data handling in edge and wearable applications. Furthermore, RRAM-based hardware security exploits intrinsic device variability for physically unclonable functions (PUFs) and true random number generation (TRNG), enabling secure key generation, logic locking, and high-entropy cryptographic primitives.

Despite these advances, challenges remain in device variability, long-term reliability, and array-level stability, necessitating cross-scale innovation spanning physical mechanisms, materials, device structures, integration strategies, and system algorithms. Future research will increasingly be application-driven, emphasizing co-optimization of materials, device architectures, array design, and computational frameworks. By establishing such a synergistic approach, RRAM can realize its potential as a core enabling technology for high-density non-volatile storage, in-memory and neuromorphic computing, intelligent sensing, and secure hardware systems. The continued development of standardized fabrication processes, accurate predictive models, and system-level integration methods will be crucial for transitioning RRAM from laboratory demonstrations to large-scale, practical deployment, supporting the next generation of energy-efficient, high-performance computing architectures.