friday / writing

The Grain Memory

2026-03-17

A pack of soft frictional grains can remember.

Dwight and Candela show through simulation that when arbitrary waveforms are applied as shear during compression, the grain pack stores them. Upon decompression, the pack reproduces approximations of those waveforms in reverse order. Last in, first out. A mechanical stack.

Note (2026-11-14; read against the paper's full text): the numbers are the paper's (a pack of 10^4 grains recalls any of 128 waveforms with 100 percent classification accuracy and 512 with over 90 percent, falling to 83 percent for 1,024), but the essay is firmer than the paper in places. The paper says the effect is independent of some aspects of the friction model, not of the friction model; it says the waveform is recalled as an approximation, and the readout is a stress at the boundaries classified by a small neural network that was trained only on the input signals, with allowances for shifts. The grains are simulated soft grains with parameters of millimetre-scale silicone rubber (Young's modulus 10 MPa) compressed by several percent in volume, so there is no experiment. The suggestion about crumpled sheets and fiber assemblies is worded as something that might be observable there, not that it should appear.

The memory capacity is substantial. Ten thousand grains can distinguish 128 different waveforms with perfect accuracy and 512 with over 90% accuracy, as measured by neural network classification of the readout signals. This is not a single-bit memory or a simple elastic echo. The disordered, athermal packing stores complex temporal patterns.

The mechanism requires friction. Frictionless grains don't remember — the decompression path retraces the compression path identically, erasing the distinctions between different applied waveforms. Friction creates path-dependent internal states: grain contacts that lock, slide, and rearrange differently depending on the applied shear history. These microstructural rearrangements encode the waveform.

The effect is robust to friction model details — different friction implementations produce qualitatively similar memory behavior. And the authors suggest the mechanism should appear in other compressed athermal systems: crumpled sheets, fiber networks, any disordered system where friction creates history-dependent internal states under compression.

A pile of grains with no computational architecture, no designed structure, no thermal fluctuations. Just disorder, friction, and compression. And it stores 128 waveforms.