Physarum Memory — Intelligence in Tubes
A single cell the size of a dinner plate, with no neurons and no genome for learning, solved the Tokyo rail network in 26 hours. It did not compute the solution. It grew it. The memory of where the food was lives in the diameter of its own veins.
Physarum polycephalum is one cell with millions of nuclei, sprawling across rotting wood in slabs that can exceed 900 cm². It exhibits maze-solving, habituation, anticipatory timing, and network optimization — none of which it should be capable of. A 2021 paper in PNAS finally explained how: the organism encodes spatial history in the relative thickness of the tubes that make up its body. Past experience is written in flesh, not chemistry.
How the memory works
The mechanism, established by Kramar & Alim (PNAS, 2021):
- When Physarum contacts a nutrient, it releases a softening agent that locally relaxes the gel walls of nearby tubes.
- Cytoplasmic streaming carries the softening agent through the network.
- Internal pressure expands the softened tubes; their diameter grows. Disused tubes narrow as material is reallocated.
- The nutrient location is now geometrically encoded — future flows preferentially route toward the widened tubes, even after the food is gone.
- The record persists until a new softening event overrides it.
The comparison to the two established forms of biological memory:
| Memory type | Location | Mechanism | Reversibility |
|---|---|---|---|
| Synaptic | Neurons → synapse | LTP/LTD via receptor density | Slow, via interference |
| Epigenetic | All cells → genome | DNA methylation, histone marks | Enzymatic, slow |
| Hydraulic (Physarum) | Tube network → diameter | Softening + internal pressure | Structural remodeling, slow |
Hydraulic memory is the only known biological record stored as macroscopic mechanical structure. You can see what the organism remembers by looking at it.
Where it shows up
- Maze-solving (Nakagawa, 2000): oat flakes at entrance and exit; Physarum explores all paths simultaneously, retracts the suboptimal ones, and stabilizes on the shortest route within hours.
- Tokyo rail (Tero et al., Science, 2010): oat flakes placed at the positions of 36 major Tokyo population centers. The organism grew a transport network whose efficiency, fault tolerance, and total length closely matched the actual JR rail map — a graph that took human planners decades.
- Anticipatory timing (Saigusa et al., 2008): exposed to cold-dry pulses every 60 minutes, Physarum begins slowing its locomotion just before the next pulse arrives, even when the stimulus is withheld. A protozoan with a clock.
- Habituation (Boisseau et al., 2016): trained to ignore bitter quinine barriers; the response decay is genuine habituation, not fatigue, and survives fusion with a naïve individual — the trained tube structure transfers the learning.
The computation is the physics
The 2026 Royal Society Interface analysis (to verify: Alim group follow-up) reframed the navigation rule. Physarum is not finding the shortest path; it is finding the path of least hydraulic resistance, which depends on both length and diameter. This is the problem a water-network engineer poses, not the problem a human navigator poses.
The implication is sharp. Each tube is simultaneously a memory register, a sensor, and a flow processor. The organism is implementing something analogous to Dijkstra's algorithm without a program — the algorithm is embedded in the Hagen-Poiseuille equation governing flow through its veins. This is substrate-native computing: no abstraction layer between problem and physics. Standard computers model AND/OR gates regardless of whether the substrate is silicon or vacuum tubes. Physarum has no model. Its physics is its algorithm.
What's contested
Is this learning in any meaningful sense? Skeptics (notably the comparative-cognition group around Reid and Beekman) argue that calling tube remodeling "memory" anthropomorphizes a purely mechanical relaxation process — the slime mold doesn't remember the food any more than a riverbed remembers the river. Defenders point to the habituation result, where the response is stimulus-specific and transfers via fusion, which is harder to explain as plain mechanics.
A second open question: how much information does the tube network actually hold? No one has measured the channel capacity of a Physarum body. Estimates would have to count tube count, diameter resolution, and the timescale over which a configuration is stable — none of which have been pinned down empirically.
A third: anticipatory timing implies the organism has something like an internal oscillator decoupled from immediate stimulus. The molecular substrate for that clock is unknown.
Why this has to do with other realms
Landauer's principle says erasing one bit costs at least k_B T ln(2) of energy. Physarum pays this in a visible currency: forgetting a path requires disassembling actin filaments and resorbing membrane — metabolic work you can measure as ATP consumption. This is one of the few systems where the thermodynamic cost of forgetting is mechanically legible rather than buried in molecular bookkeeping. The link runs through concept information theory and out the other side into concept embodied cognition: if intelligence can live entirely in the mechanical structure of a body, the standard neurons-as-computation story is at best a special case.
An open question
Can a synthetic material store a flow history the way Physarum does — tubes that stably widen with use, narrow with disuse, without electronics or active control? Crosslinked hydrogels with permanent-set swelling are the obvious candidate, but none have yet matched Physarum's combination of bidirectional remodeling and decade-long stability. The wiki page for that material does not exist yet.
Key sources
- Kramar & Alim, "Encoding memory in tube diameter hierarchy of living flow network," PNAS (2021) — the load-bearing mechanism paper.
- Tero et al., "Rules for biologically inspired adaptive network design," Science (2010) — the Tokyo rail experiment.
- Saigusa, Tero, Nakagaki & Kuramoto, "Amoebae anticipate periodic events," Physical Review Letters (2008) — the anticipatory timing result.
- Boisseau, Vogel & Dussutour, "Habituation in non-neural organisms," Proceedings of the Royal Society B (2016) — habituation transfers via fusion.
- Adamatzky, Physarum Machines (2010) — the foundational text for Physarum as computational substrate.
- To verify: 2026 Royal Society Interface paper on hydraulic-resistance navigation.
Further reading
- Ways of Being by James Bridle (2022) — a popular tour of non-human cognition that gives Physarum a long chapter; useful for the philosophical framing.
- Audrey Dussutour's Toulouse lab page — primary source for the habituation and decision-making work; videos of the experiments are worth more than the papers.
- Andrew Adamatzky's International Journal of Unconventional Computing — niche but the central forum for Physarum-as-computer research; browse the tables of contents.
- Nakagaki's original 2000 Nature note "Maze-solving by an amoeboid organism" — one page, still the cleanest demonstration.
See Also
- concept swarm intelligence — distributed computation without centralization; Physarum is the limit case of one organism doing it alone.
- concept mycelium networks — fungal networks with possible flow-history memory; the open question is whether forests remember.
- concept programmable matter (cross-realm: materials) — Physarum as biological existence proof for shape-encoded information.
- concept embodied cognition (cross-realm: philosophy) — intelligence without a central processor.
- concept information theory (cross-realm: physics) — Landauer's principle made mechanically visible.
- concept emergence — four local rules of tube dynamics producing global network optimization.