Memory is not computation until the system closes the loop itself

What the work claims

The author claims that memory, feedback, and rich state-dependent dynamics do not by themselves establish computation, and proposes a precise closure criterion that separates physical systems that merely store and transform state from those that compute autonomously.¹ Computation is defined as robust coarse-grained transition preservation: a physical map implements an abstract transition only when a coarse-graining of its state space satisfies compositionality, with abstract states realized by separated physical basins and transitions stable under noise. Autonomous computation adds a closure condition: an internal physical readout state must select the next physical operation, in the form x(n+1) = F(C(y(n)))(x(n)), where y is read out from the physical state and C maps that readout to the operation to perform.

Applied to the wave-particle walker of Perrard, Fort, and Couder, the criterion delivers a nuanced verdict: the walker is a wave-memory machine with genuine Turing-like primitives (writing, finite-time storage, local reading, feedback, finite-time reversal, and externally triggered erasure), but not a closed autonomous physical computer, because the phase shift that erases the wave memory is imposed by the experimenter. The framework then converts the distinction into a design principle: memory becomes autonomous computation when physical readout basins are coupled back to operation selection.

This is a theory paper: a single-author, mathematics-first argument with one motivating physical case study and a constructive section proposing extensions (multistable readouts, thresholded wave detectors, boundary-controlled reservoirs, coupled walkers, and a minimal autonomous eraser) that would move the walker across the line.

How it works

The test case is the hydrodynamic quantum analogue. A droplet bounces on a vertically vibrated fluid bath near the Faraday threshold; each impact excites standing waves that decay slowly, so the present wave field holds a fading trace of past impact positions; the droplet is propelled by the local slope of its own wave field. The author formalizes this as a stroboscopic reservoir with state x(n) = (position, velocity, phase, wave field), where the wave field stores an exponentially decaying trace of previous impacts and guides future motion through local slope coupling. The model separates the ingredients that matter for the definition: physical writing, storage, reading, feedback, and externally triggered erasure.

The definition builds in two stages. First, coarse-grained transition preservation: abstract symbols correspond to basins in the physical state space that are separated enough to be distinguished at finite resolution, and the physical dynamics must carry basins to basins in a way that composes like the abstract transition system and survives noise. This is what distinguishes computation from mere dynamics: a rock falling does not compute its trajectory because its microscopic state does not implement any robust, composable, symbol-preserving map. Second, closure: the readout is not merely observed by an experimenter but is coupled back into the dynamics so that the value of the internal readout state determines which physical operation the system performs next. A system whose outputs are read, interpreted, and acted upon by an external agent fails this condition at exactly the point where the agent intervenes.

The walker passes the first stage and fails the second: its erasure is a pi phase shift applied by the experimenter, so the memory loop is not closed within the system. The appendices work the same analysis for trajectory memory, wave-energy bookkeeping, and failure modes (backtracking and wave cancellation after the phase intervention), and state the classification as a formal proposition.

Where a skeptic should push

The most load-bearing assumption is that closure, as defined, is the right line to draw. Definitions of physical computation are contested territory, and reasonable people defend pancomputational, counterfactual, and semantic alternatives. The criterion is convincing for what it excludes, but exclusions are easier than validations: showing that a readout is coupled back does not by itself show the coupled system computes anything useful, only that it is autonomous in the technical sense.

Second, the classification of real experimental systems will often hinge on where the system boundary is drawn. Call the shaker, the bath, and the droplet one system and erasure is internal; carve the experimenter's phase control into the boundary and it is external. The paper handles this carefully for the walker, but practitioners will be tempted to redraw boundaries until their favorite device passes, and the framework needs discipline against that.

Third, the empirical content is one test case plus proposed extensions that were not built. The constructive section sketches what a closed walker would look like and even outlines an experimental validation, but as written, no demonstrated physical system in the paper satisfies the full closure criterion. That is honest, but it means the framework's practical yield is currently a design checklist, not a catalogue of autonomous machines.

A closure test for claims of organoid computation

This paper is the most directly relevant theory organoid intelligence has been handed in a while, because its verdict structure maps one-to-one onto the field's standard experiment. A typical organoid computing demonstration looks like this: electrodes record spontaneous or evoked activity, an external computer trains a readout or selects stimuli, a stimulator writes the chosen pattern back, and the culture is credited with computing. Under the closure criterion, that system is a reservoir with an external operator, in exactly the walker's position: writing, storage, reading, and feedback occur, but the operation selection (which stimulus next, when to reset, what counts as the trial boundary) is imposed by the experimenter's code. By this criterion, current organoid rigs are wave-memory machines, not closed computers, and that is a precise statement rather than an insult.

The non-obvious implication is architectural, and it sharpens the field's research program rather than deflating it. Closing the loop means the tissue's own state must select the next physical operation: activity-dependent stimulation protocols, where the detected pattern itself triggers the write-back rather than a policy running on a PC, are steps in this direction. The paper's design principle (couple physical readout basins back to operation selection) reads, in organoid terms, as: build the trial logic into the interface so that the culture's electrophysiological state gates what happens to it next. A system that does that would be genuinely novel in the OI literature, and it would make a stronger claim to autonomy than any accuracy benchmark has.

There is also an epistemic threat the paper states better than the OI field has. One of its discussion sections warns against premature computational interpretation of cortical traveling waves: waves propagate, modulate excitability, organize spike timing, and may participate in neural computation, but wave presence is not evidence of computation. Organoid electrophysiology is full of narrative in which bursts, oscillations, and network waves are described as the organoid computing. The closure criterion supplies a check: which transitions are preserved under a coarse-graining, which basins are separated, and what, if anything, is closed? Asking those questions would retire a large fraction of the field's loosest claims. The dual-use and ethics angle cuts the same way in reverse: a system whose operation selection is genuinely closed inside living tissue is harder to monitor, harder to interrupt, and harder to attribute responsibility for, which raises governance questions that current OI oversight discussions have not seriously started.

The bottom line

Established: a rigorous two-stage definition of physical computation (robust coarse-grained transition preservation plus readout-to-operation closure), a careful stroboscopic model of the wave-particle walker, and a formal classification of that system as a wave-memory machine with Turing-like primitives but externally imposed erasure. Hypothesis, not established: that the criterion will be accepted as the standard boundary, that any existing physical device currently satisfies full closure, and that the proposed walker extensions behave as sketched.

For organoid intelligence, the paper's gift is a test. What would confirm its relevance is the field adopting the closure language and producing at least one stimulation system in which the culture's own detected state selects the next operation; what would weaken it is the boundary-drawing problem proving so slippery that no experimental claim can be cleanly classified. Either outcome would still leave the field with better questions than it has now.