Memristors
based on two-dimensional (2D) materials can exhibit
great scalability and ultralow power consumption, yet the structural
and thickness inhomogeneity of ultrathin electrolytes lowers the production
yield and reliability of devices. Here, we report that the self-limiting
amorphous SiO
x
(∼2.7 nm) provides
a perfect atomically thin electrolyte with high uniformity, featuring
a record high production yield. With the guidance of physical modeling,
we reveal that the atomic thickness of SiO
x
enables anomalous resistive switching with a transition to an analog
quasi-reset mode, where the filament stability can be further enhanced
using Ag–Au nanocomposite electrodes. Such a picojoule memristor
shows record low switching variabilities (C2C and D2D variation down
to 1.1 and 2.6%, respectively), good retention at a few microsiemens,
and high conductance-updating linearity, constituting key metrics
for analog neural networks. In addition, the stable high-resistance
state is found to be an excellent source for true random numbers of
Gaussian distribution. This work opens up opportunities in mass production
of Si-compatible memristors for ultradense neuromorphic and security
hardware.