Memristors with tunable conductance characteristics have attracted great attention in high density memory and neuromorphic computing. However, the dynamics of conductance change for filamentary-type memristors is generally asymmetric: The set transition is quite abrupt, while the reset transition is usually gradual, which is a big challenge to achieve continuous conductance tuning characteristics in both set and reset processes. In this work, we demonstrated an HfO2/TiOx (10 nm/10 nm) bilayer structure memristor with the feature of bidirectional conductance tuning (a gradual increase or decrease in conductance) in a simple pulse-train operation mode. A series of voltage pulses with specific amplitude and a fixed width of 50 ns were used to realize the characteristics of bidirectional conductance tuning. By further optimizing the pulse amplitude conditions, such as −1.1 V/50 ns for the set process and 1.3–1.4 V/50 ns for the reset process, the conductance of the memristor can be tuned almost linearly with the input pulse voltage. Such linear conductance update is highly desired for improving the fault tolerance ability in massive data storage or neuromorphic computing.
Lead-free Cs3Bi2Br9 perovskite memristors have emerged as promising candidates for nonvolatile memory and neuromorphic computing due to their environmental friendliness and stability. However, their practical applications are hindered by severe endurance degradation during repetitive resistive switching (RS) cycles, attributed to incomplete RESET of conductive filaments (CFs) at grain boundaries (GBs). Herein, we systematically investigate this degradation mechanism and propose two innovative strategies to enhance device endurance: (1) grain engineering via thermal annealing to reduce GB density and (2) GB passivation using polyethylene glycol (PEG) additives. By optimizing annealing temperatures (100–250 °C), Cs3Bi2Br9 films exhibit progressively larger grain sizes (158.6–295.2 nm), which delays high resistance state degradation by minimizing CF nucleation sites. However, residual GBs still permit partial Ag accumulation. To address this, PEG incorporation chemically passivates GBs, effectively blocking ion migration and trapping. The optimized PEG-passivated Ag/Cs3Bi2Br9:PEG/ITO memristor achieves remarkable improvements: endurance is extended from 100 to 1000 cycles, and the switching window expands from 6 to 176 times. Furthermore, the optimized device demonstrates continuous conductance modulation for multilevel storage and synaptic function emulation, enabling a 784–100–10 fully connected neural network to achieve 92.2% recognition accuracy on the MNIST dataset after 100 training epochs. This study elucidates the GB-mediated degradation mechanism and demonstrates that combining grain size optimization with GB passivation provides a universal framework for high-performance RS devices in neuromorphic computing and nonvolatile memory.