A memristor that forgets on command, and what it teaches the electrode engineer

What the work claims

The authors build two-terminal vertical devices from exfoliated CuCrP2S6 (CCPS), a van der Waals layered material, sandwiched between a Ti(2 nm)/Au(10 nm) bottom electrode and a Cr(5 nm)/Au(50 nm) top electrode. They show robust resistive switching with no measurable ferroelectricity, ruling out the polarisation mechanism that dominates most of the thiophosphate literature by direct piezoresponse force microscopy. The switching, they argue, is purely ionic: native Cu+ ions hop across the van der Waals gaps under field and undergo quasi-reversible reduction and re-oxidation at the electrode interfaces.¹

The bold move is the second claim: the decay time constant of the written conductance is itself programmable. A fast Set sweep leaves a state that relaxes back quickly; a slow Set sweep leaves one that persists. Volatility becomes a dial, not a defect. This is a primary experimental result on a handful of single devices, not a validated technology; weight it accordingly.

How it works

The devices switch with triangular sweeps of order seconds: 1.13 V/s corresponds to a roughly 5 s pulse and produces almost fully volatile conductance change, while 0.25 V/s, roughly 24 s, produces a remanent change that is still only partially relaxed after a 5 minute delay. Readout at 1 V shows an On/Off ratio tunable from 2 at a 5.67 V/s programming sweep to 16 at 0.15 V/s. The relaxation of the readout current fits a stretched exponential with a disorder exponent beta of about 0.5 to 0.6, the signature of a broad distribution of relaxation pathways, consistent with CCPS behaving as a disordered solid electrolyte.¹

The proposed mechanism couples two processes. Fast processes are Cu+ redistribution across the van der Waals gaps, which screens or unscreens internal bias fields and modulates interfacial injection barriers; slow processes are electrochemical, namely the interfacial reduction Cu+ + e- to Cu0 and its reverse. Thresholds sit at Reset of -1.5 V or below and Set of 2 V or above, saturating beyond about 3.5 V, corresponding to a field of roughly 0.045 to 0.06 V/nm across a tens-of-nanometres flake, on the order of the field at which ionic migration dominates in the sister material CuInP2S6. Kelvin probe force microscopy on a 17 to 30 nm flake in argon shows written surface-potential patterns only above about 1 V, with a 200 mV contrast between regions poled at -6.5 V and +5 V.¹

The electrochemical half is nailed down quantitatively. In a soft-breakdown regime, the devices show bipolar transient peaks whose height scales as the square root of sweep rate with a zero intercept, exactly the Randles-Sevcik signature of a freely diffusing, quasi-reversible redox species rather than a surface-bound one. Fitting that relation gives a Cu+ diffusion coefficient of 3.65 x 10^-12 cm2/s, in good agreement with 5.00 x 10^-12 cm2/s derived from published ionic-conductivity data for CCPS, and well above the 2.1 x 10^-13 cm2/s similarly derived for CuInP2S6. Pushed harder, the same redox forms metallic filaments that are themselves volatile, dissolving on the ramp down, and overcycling eventually leaves irreversible damage: Raman spectra pick up the S-S stretching modes of amorphous sulfur where Cu+ has been extracted from the lattice. A window of genuinely reversible operation exists below that damage regime, and low-voltage cycling can even redissolve incipient filaments.¹

Where a skeptic should push

The load-bearing assumption is that the switching observed at second-scale sweeps is the same physics that would survive in a useful memory element. The evidence is six devices, with the rate and volatility trends confirmed on three; every electrical number above comes from micron-scale exfoliated flakes measured one at a time, several under 0.12 mbar of air or in argon, not in arrays, not in ambient saline-equivalent conditions, and with no endurance or retention statistics beyond the traces shown. The model is explicitly qualitative: the authors state they cannot tell which interfacial barrier limits conduction, and that the limiting barrier may change during switching.

Second, the tunable volatility is demonstrated at tens of seconds. That is a long way from the millisecond to second synaptic timescales neuromorphic marketing prefers, and the paper's own prior-art framing notes earlier vertical CCPS devices operated below 1.2 V for back-end-of-line compatibility, whereas the interesting redox phenomenology here lives at 2 V and above. There is also a selection hazard: devices that showed soft breakdown were pursued for the redox analysis, so the prevalence of the filamentary regime across a population is unknown. Treat this as careful mechanism work on a promising material, not as a device benchmark.

What tunable volatility means for MEA electrodes

The non-obvious implication runs in both directions, and neither is comfortable.

As an opportunity: the acquisition chain around a microelectrode array is full of places where a programmable, self-erasing conductance would do real work. Baseline wander, stimulation residuals and slow impedance drift on chronic electrodes are currently fought with digital high-pass filters, re-referencing and periodic recalibration, each of which distorts the very low-frequency and spike-adjacent content the array exists to capture. A two-terminal element whose leakage time constant is set by how it was last written, in the seconds-to-minutes regime this paper demonstrates, is the hardware primitive for that job: an analog, adaptive, power-free relaxation stage that tracks the interface instead of being hand-tuned against it. The paper's central engineering lesson is that you get that programmability for free from ionic relaxation, provided you operate inside the reversible window and stay clear of filament formation.

As a threat: this paper is the cleanest recent demonstration that the electrode-electrolyte junction of an MEA is not passive, and is not static on any timescale the acquisition chain cares about. Mobile ions redistribute under fields of a few hundredths of a V/nm; interfacial redox at noble-metal electrodes is quasi-reversible at room temperature; written charge states screen themselves over minutes with stretched-exponential kinetics. Every one of those sentences describes failure modes MEA operators measure as impedance drift, polarization offsets after stimulation, and reference-electrode wander. The practical transfer is diagnostic: the Randles-Sevcik test the authors use, checking whether a transient current scales with the square root of sweep rate, is a bench procedure you can run on an electrode to detect freely mobile ionic species contaminating the interface, quantifying the drift instead of merely observing it. A further caution follows from the irreversible sulfur Raman signature: an electrode driven repeatedly into over-voltage is being electrochemically rebuilt, quietly, and the rebuild does not reverse.

The bottom line

Established: in exfoliated paraelectric CCPS, resistive switching can be purely ionic, its volatility can be tuned over an order of magnitude by programming sweep rate, and interfacial Cu+ redox is confirmed quantitatively by a diffusion-coefficient extraction. Hypothesis, still: that this platform scales to dense, CMOS-integrated, long-endurance arrays; nothing here tests that. What would confirm the capability story is retention statistics across a device population and volatility control at millisecond write timescales; what would break it is evidence that the reversible window collapses in physiological electrolytes at 37 C. For the instrumentation reader the paper is worth attention either way, because it writes down, with numbers, exactly how an ionic interface remembers and forgets, and that interface is the first component of every acquisition chain this field builds.