Lipid-membrane memristors and what living tissue must still provide

What the work claims

This is a grant, a funded plan and not a benchmarked result, so it must be read as intention supported by a published track record rather than as evidence in itself.¹ It is a Foundations of Emerging Technologies Small award of 336,655 dollars obligated in fiscal 2024, running to April 2027, to build physical reservoir computers out of what the award calls ion-channel-based memristors. The central claim has two parts. First, that an insulating lipid membrane hosting voltage-activated ion channels exhibits voltage-dependent memristance, a resistance that depends on the history of applied voltage, and that these devices carry collocated volatile and non-volatile memory, a fading short-term component and a slower persistent one in the same element. I flag that the durable non-volatile state is the novel and least-established part of this claim, because most published lipid-membrane memristors in this lineage are short-term, volatile devices, and the single-layer thesis below rests on the persistent component, so it inherits that uncertainty. Second, and more pointed, that because these devices carry collocated volatile and non-volatile memory, the reservoir layer and the trained readout layer of a reservoir computer might be combined into a single physical layer, rather than kept as separate stages. The award record lists four associated 2025 publications, three in journals (including npj Unconventional Computing and Nanoscale) and one in an IEEE conference proceedings, which establishes this as an active, publishing line of work rather than a proposal on paper.

How it works

Start with the device. A memristor is a two-terminal element whose resistance depends on the history of current or voltage through it, so it stores information in its own conductance. The unusual choice here is to make one out of biology's own parts. The basic element is a synthetic lipid bilayer, the same two-molecule-thick oily film that forms every cell membrane, assembled between two droplets or across an aperture. On its own it is an insulator. Add voltage-activated ion channels, protein pores that open and close depending on the voltage across the membrane, and the film gains a voltage-dependent conductance with memory: the channel population and the membrane's own dynamics respond to recent voltage and relax on their own timescales. That gives exactly the ingredient a reservoir needs, a nonlinear response with fading memory, built from materials chemically identical to a real membrane rather than from silicon or metal oxides.

Now the computing frame. Reservoir computing takes a fixed, high-dimensional nonlinear system, drives it with an input, and trains only a simple linear readout of the system's state. The internal connections are never trained, which is what lets the approach exploit messy physical media that no one can rewire. Traditionally the reservoir and the readout are separate, the reservoir supplying rich dynamics and the readout supplying trained, non-volatile weights. This award's sharpest technical bet is that a device holding both volatile and non-volatile memory at once can be its own readout, collapsing two layers into one. The plan is to validate the idea on crossbar arrays of these membranes, and, tellingly, to generalize the resulting architecture so it also works with other volatile and non-volatile memristors, possibly solid-state ones.

Where a skeptic should push

The category caution comes first: this is a Small award whose purpose is to establish feasibility, and an abstract carries no benchmarks, no task accuracy, no throughput, no device-to-device variability. The load-bearing assumption worth stressing is durability and scale. A lipid bilayer is a fragile object, sensitive to temperature, mechanical disturbance, and time, measured in a laboratory rig rather than packaged as a chip, and a crossbar of many such membranes inherits every one of those fragilities plus the problem of making them uniform. The single-layer claim, reservoir and readout in one device, is elegant but unproven at the abstract stage, and combining two functions in one physical element often trades away the independent control that made each function reliable. Separate the demonstrated from the asserted: the award asserts memristance and proposes the reservoir architecture, and the listed publications suggest device-level results exist, but I did not read them and cannot vouch for the system-level performance, which is where reservoir schemes usually succeed or fail.

When molecules undercut the living reservoir

Organoid intelligence usually argues from two premises: that biological materials compute with an efficiency and richness silicon cannot match, and that being made of the right stuff, membranes, ion channels, real neural dynamics, is part of the advantage. This award is a wetware project that quietly attacks the second premise from inside the tent. An ion-channel memristor is biological in material, a genuine lipid membrane with genuine ion channels, yet it is not alive, does not need feeding, does not develop, and can in principle be fabricated to spec on a crossbar. If such a device delivers the nonlinear fading-memory kernel that reservoir computing needs, then it captures much of the biological-computing pitch, the low-power analog dynamics of membranes and channels, without the culture, contamination, vascularization, and month-long maturation that make organoids hard. The non-obvious implication is that the living organoid is not competing mainly against silicon here, it is competing against a cheaper, more controllable piece of biology, and on the reservoir task specifically the molecular device would hold the deployment advantages if it matches organoid or silicon reservoirs on a real benchmark, which the award has not yet shown.

The threat is crystallized by the award's own fourth aim: to generalize the architecture so it runs on other memristors, possibly solid-state. That is at least consistent with the computational value living in the device dynamics, the collocated volatile and non-volatile memory, rather than in the biology per se, though a generalization aim is also routine engineering ambition and does not by itself prove the biology is dispensable. If the scheme did port to solid-state, it would suggest the membrane was a convenient host for a dynamical property rather than an irreplaceable source of it, and by extension the living organoid, whose reservoir role rests on the same fading-memory property, would be exposed to a similar substitution. This is the substrate-obsolescence argument arriving from an unexpected direction, not from a silicon neuromorphic chip but from a fellow traveler in wetware that is easier to build.

There is a real opportunity in the same facts, and it is where the durable case for living tissue actually sits. Reservoir computing deliberately freezes the reservoir and trains only the readout, which means it throws away the one thing living neural tissue offers that a fixed lipid-membrane reservoir does not: ongoing, activity-dependent synaptic plasticity, the capacity of the substrate to rewrite its own internal weights through use. One caveat keeps this honest, and it matters: neuromorphic silicon already implements local, autonomous learning rules such as spike-timing-dependent plasticity, so plasticity as such is not unique to living tissue. The defensible remainder is narrower and empirical, that biological plasticity rules, running in dense three-dimensional self-developed recurrence, might outperform the plasticity a chip can be programmed to run, which is an unproven bet rather than an established advantage. An ion-channel memristor has memory but not learning in that sense; its non-volatile state is set by applied voltage, not by a biological plasticity rule the tissue runs itself. So the ICM result is clarifying rather than purely threatening. It says that if the field's case for organoids rests on being a good fixed reservoir, that case is weak and shrinking, because cheaper biology and cheaper silicon can both supply a fixed nonlinear kernel; and it says the defensible case for living tissue is the part reservoir computing ignores, but narrowed by the caveat above: not plasticity in the abstract, which silicon also has, but the open, currently unevidenced possibility that biological plasticity rules and dense self-developed three-dimensional recurrence furnish something an engineered reservoir or an on-chip learning rule cannot cheaply match. The ICM device is also a plausible ally: because it is built from the same biomolecular materials as a cell membrane, it could in principle interface with or sit alongside living tissue rather than only competing with it, though being made of lipid is not the same as being implant-safe.

The bottom line

Established as facts of the award: NSF funded a Penn State effort to build physical reservoir computers from ion-channel memristors, lipid membranes with voltage-activated channels that show voltage-dependent memristance and collocated volatile and non-volatile memory, with a stated aim of collapsing reservoir and readout into one layer and of generalizing to other memristors including solid-state ones, and the principal investigator has a 2025 publication record in the area. Everything about task performance, durability, and scale is beyond the award record, and I did not read the listed papers, so my reading is bounded to the plan and to reservoir-computing fundamentals. For organoid intelligence the calibrated conclusion is that this is a wetware competitor, not an ally by default: on the fixed-reservoir task a molecular device can plausibly match the living substrate with far less overhead, which narrows the organoid's unique value to what a frozen reservoir cannot do, chiefly self-modifying plasticity and possibly a richer recurrent kernel. The threat would be confirmed if an ion-channel-memristor reservoir matches organoid or silicon reservoirs on a real temporal benchmark, and the fourth aim's port to solid-state succeeds; the living substrate's case would be vindicated only where in-tissue plasticity delivers learning that no fixed reservoir, biological or otherwise, can reproduce.