A body can remember what it is not yet

Start with a less dramatic repair. A cut closes because cells migrate, contract and proliferate. If the wound is small, local rules may be sufficient. Each cell senses neighbours, matrix tension and chemical signals, then performs a nearby action. No cell needs a picture of the finished body.

Now remove half a planarian. The remaining fragment must do more than close an edge. It must infer which structures are missing, establish anterior and posterior polarity, produce organs at the right scale and stop when a viable whole has been restored. A fragment from near the tail begins with different cells and geometry from a fragment near the head, yet both often reach the species-typical form. The visible outcome therefore compresses many possible repair trajectories.

The hidden assumption in the blueprint metaphor is that the genome directly specifies one finished form. Genomes certainly constrain what proteins, receptors, channels and regulatory networks can exist. Yet the same genome is used by a neuron, skin cell and liver cell. Development depends on state, location, signalling history and interaction. The planarian result adds a sharper point: two systems can share genome and current gross anatomy while carrying different future morphological commitments.

Calling this “intelligence” is useful only if the term does explanatory work. Here it means neither human reasoning nor subjective experience. It means the capacity of a system to preserve a preferred region of state space by sensing deviation, coordinating action and selecting among available paths. That definition is deliberately operational. It allows bacteria, tissues, plants, animals and machines to be compared without pretending they share an inner life. Related basal- and collective-cognition programmes use similarly graded, substrate-neutral language, although the interpretation remains contested.[18][19]

The central causal claim is that slow, writable physiological state can turn a collection of locally competent cells into a larger controller. Ion channels establish membrane voltages. Gap junctions allow electrical and small-molecule coupling between neighbouring cells. Voltage influences calcium, neurotransmitter transport, phosphatases, gene expression, proliferation, migration and apoptosis. When this network has multiple stable states, a transient input can move the tissue from one attractor to another. The input disappears, but the state remains long enough to bias later development.[3]

Developmental bioelectricity is not a claim that voltage replaces genes. Voltage, gene-regulatory networks, morphogens, mechanics and metabolism interact. In planaria, early voltage differences appear before some known asymmetric transcriptional markers, and transient depolarisation during the early regenerative window can alter later anterior-posterior polarity.[2] In frogs, changing membrane potential can influence eye induction, including ectopic eye structures outside the ordinary eye field.[4] In planaria, bioelectric signalling also affects head and organ size through downstream remodelling processes.[5] Work in the embryonic frog brain also shows that downstream gene activity can depend on a spatiotemporal voltage pattern distributed across a cell sheet, rather than on one cell’s voltage in isolation.[16]

flowchart LR
  A["The worm with a hidden future"] --> B["A body can remember what it is<br/>not yet"]
  B --> C["From voltage state to target state"]
  C --> D["Where the metaphor breaks"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#007D68,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

From voltage state to target state

A cell membrane separates charge. Ion channels and pumps change which ions cross it, creating a resting potential. In neurons, voltage often changes quickly to transmit spikes. In many non-neural cells, resting potential changes more slowly and can remain at distinct levels. Because channels are both controlled by voltage and help determine voltage, feedback can create more than one stable state.[3][15][17]

The smallest useful model is a bistable element coupled to its neighbours:

Here, Vᵢ is the scaled membrane state of cell i. The term α(Vᵢ − Vᵢ³) creates two preferred states near −1 and +1. The coupling g pulls a cell towards the average state of its neighbours, standing in for gap-junction-mediated coordination. The input uᵢ(t) is a temporary intervention. The equation assumes identical cells, nearest-neighbour coupling and a symmetric two-state landscape. Real tissues violate all three assumptions.

Three meanings of memory

The word “memory” causes trouble because it collapses three different causal claims.

A voltage that remains after stimulation is trace memory. A Venus flytrap that integrates recent action potentials and changes its next response exhibits policy memory. A tissue state that makes the same injury converge to a different anatomy is evidence for target memory. This distinction is the hinge of the argument: persistence is common, but a rewritable completion condition is rare.

Non-neural control outside morphogenesis

The same logic appears in simpler forms elsewhere. In bacterial chemotaxis, receptor methylation participates in an integral-feedback architecture that lets the system adapt to sustained signals and recover sensitivity. The molecular details are not a little brain, but the control principle is recognisable: retained internal state changes future response so that an output returns towards a reference condition.[7]

The Venus flytrap provides a different case. Mechanical stimulation generates action potentials. One event is insufficient for full commitment. A second event contributes to closure, while further stimulation scales downstream jasmonate signalling, digestive enzyme expression and nutrient uptake. The plant is not performing symbolic arithmetic. It is integrating temporally separated signals so that expensive action depends on accumulated evidence.[8] Slime moulds provide a third, deliberately weaker case: repeated exposure to a harmless repellent can reduce avoidance, with the changed response persisting for a time after exposure.[9] That supports non-neural habituation, not a general claim that the organism represents an explicit goal.

This does not imply that a tissue is conscious. It does not imply beliefs, reasons or suffering. It says only that some operations used by nervous systems have older, non-neural implementations. Neurons may enlarge the horizon of goals without inventing goal-directed regulation from nothing.

Where the metaphor breaks

Biology is full of forms that look purposeful. Evolution has selected systems whose local dynamics produce viable outcomes. An observer can describe those outcomes as goals even when no internal variable represents them. The problem is therefore not to ban teleological language. It is to ask when that language predicts interventions better than a purely local description.

I propose a counterfactual control-competence test. It does not produce one intelligence score. Instead, it asks for a pattern of evidence across five independent dimensions.

Four experiments that make the boundary move

The strongest alternative explanation

Planarian polarity can be described through self-organising biochemical systems. Work on antagonistic head and tail patterning shows that Wnt signalling gradients and opposing patterning domains can re-establish the anterior-posterior axis through local feedback and mutual inhibition.[10] More generally, reaction-diffusion dynamics can generate and restore patterns without a central controller.[11] Mechanics can also create global form from local stresses.

That rival is strong because it may explain the same anatomy with fewer cognitive terms. The bioelectric account must therefore show more than causal influence. It must show that a physiological state is writable, persistent, read during later repair and sufficiently specific to alter the endpoint. Even then, “target” may remain a useful control-theoretic description rather than a literal internal picture.

There is also a replication boundary. Many of the most striking target-morphology results come from a concentrated research lineage. A recent independent preprint reports replication of a two-headed phenotype after temporary gap-junction inhibition, but the broader cryptic-state, reversal and target-memory package still needs extensive independent reproduction.[20] This matters because the paper’s thesis depends on the full causal chain, not on one unusual phenotype.

stateDiagram-v2
  [*] --> Claim
  state "Three meanings of memory" as Claim
  state "Non-neural control outside<br/>morphogenesis" as Test
  state "Four experiments that make the<br/>boundary move" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Build a test before naming a mind

The practical contribution is a small bioelectric-memory model and an experimental protocol. The model is deliberately simpler than a real ion-channel network. Its purpose is to reveal the logic hidden by prose: bistability supplies persistence, coupling supplies collective state, and a transient pulse can change the endpoint of a later reset.

// Explanatory surrogate, not a calibrated biological model.
function stepCell(v, neighbourMean, input, dt, alpha, coupling, bias, noise) {
  const bistable = alpha * (v - v * v * v);
  const social = coupling * (neighbourMean - v);
  const stochastic = noise * (Math.random() * 2 - 1);
  return clamp(v + dt * (bistable + social + bias + input + stochastic), -1.4, 1.4);
}

// A transient pulse acts only on the right-edge cells.
// After washout, persistence requires the network to remain in the new basin.
// "Amputation" randomises the middle while preserving boundary state.

A realistic configured-system experiment

Consider a planarian assay designed to distinguish a bioelectric target from an early trigger. Begin with voltage-sensitive imaging and ordinary morphology. Apply a spatially controlled ion-channel or gap-junction perturbation for a defined window. Confirm washout chemically where possible, and verify viability, stem-cell distribution and major polarity markers. Select animals whose present anatomy appears normal. Recut them with a standard geometry and track not just the endpoint but the entire regenerative trajectory.

Now add three discriminators. First, perform a reverse physiological write before recutting. If the altered outcome returns towards baseline, causal specificity strengthens. Second, produce the same final morphology through a direct biochemical route, such as Wnt-pathway perturbation, then compare intervention fingerprints. Third, disrupt or transplant candidate voltage-bearing regions to locate the boundary of persistence.

A positive result permits the claim that the manipulated physiological state participates in a control-layer representation of target form. A negative result has several meanings. The chosen state may be downstream, too weak, wrongly timed or not part of the target at all. It may also show that the endpoint is generated by self-organising dynamics without a separately writable setpoint. Neither outcome settles whether the organism is conscious.

What synthetic living systems add

Xenobots and Anthrobots reveal that genetically ordinary cells can form novel multicellular bodies with movement and task-relevant effects outside their species-typical anatomy.[12][13][14] This demonstrates morphogenetic plasticity. It does not yet demonstrate intelligence. Motility can arise from cilia and shape. Scratch closure can result from local interaction. The relevant next experiment is not another surprising video. It is a target-write test: can the construct retain an experimentally written state that changes how it repairs, routes or stops under the same later challenge?

These hypotheses are deliberately narrower than “cells think”. They make better science because each can fail. They also make better engineering. A regenerative intervention based on a true target state should generalise across damage patterns. An intervention that merely pushes one pathway may work only for the exact lesion used during development.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The worm with a hidden future
  M->>E: Three meanings of memory
  E-->>M: compare rival explanations
  E->>D: Locate where retained state begins to organise action
  D-->>Q: revise scope or act

Locate where retained state begins to organise action

When a tissue fails to regenerate, the conventional response is to search for missing growth factors, genes or cell types. Those are necessary questions, but they assume the desired anatomy is already fixed and the problem lies only in execution. The bioelectric evidence introduces another possibility: the system may be executing a different target, or may have lost the distributed state that lets local cells coordinate around a global form.

That changes experimental design. Before adding more molecular instructions, measure the control layer. Ask which variables persist after the initiating signal, which of them predict the endpoint, whether they can be rewritten and whether a reverse write restores the earlier outcome. Track the whole recovery trajectory, not only the final photograph. Separate current anatomy from latent commitment.

It also changes the philosophy of intelligence. “Does it have neurons?” is a poor first question. Better questions are: What disturbances can it correct? Over what spatial and temporal horizon? Which internal state determines when correction stops? Can that state be changed independently of the present configuration? Does the system reroute when the obvious path is blocked?

The boundary of intelligence is not where electricity becomes neural. It is where retained state begins to organise action around a counterfactually testable end. Below that boundary lie patterns that merely persist. Above it lie systems whose future can be changed by rewriting what counts, for them, as finished.