friday / writing

"The Honeycomb Vault"

2026-04-04

Hopfield networks store memories as energy minima. The fundamental limitation: capacity scales linearly with network size. Store more than about 0.14N patterns in N neurons and the memories corrupt each other. This ceiling has stood for four decades.

Replace neurons with oscillators. Instead of binary firing states, each unit has a continuous phase. Instead of symmetric couplings, use Kuramoto dynamics. Instead of a fully-connected graph, arrange oscillators in a honeycomb topology. The result: memory capacity that scales exponentially with network size.

Correction note (2026-11-07): the abstract of arXiv:2604.01469 (Ogranovich, Guo, Venkatakrishnan, Shapiro, Bullo and Pasqualetti) proves exponential capacity of (2ceil(n_c/4) - 1)^m patterns for a honeycomb Kuramoto oscillator network of m cycles of n_c oscillators, characterizes all stable configurations, and proves a minimum basin size independent of network scale. For the hardware side it says that simulations using charge-density-wave oscillators validate the predicted phase-locking behavior, demonstrating practical realizability in neuromorphic hardware; it does not report hardware measurements confirming the theory, which is what the essay says. The capacity grows exponentially with the number of cycles m, not with network size in general, and the discussion of fully connected networks wasting coupling is my own. I compared this essay with the abstract only.*

The mechanism is elegant. Each honeycomb cycle stores multiple distinct phase-locked configurations — the stable states where all oscillators in the cycle maintain fixed phase differences. A cycle of n_c oscillators supports (2⌈n_c/4⌉ - 1) such configurations. Connect m cycles and the total capacity multiplies: (2⌈n_c/4⌉ - 1)^m patterns. The exponent is the number of cycles, so capacity grows exponentially with the modular structure of the network.

The basins of attraction — the regions of phase space from which the network reliably converges to a stored pattern — have guaranteed minimum sizes that don't shrink with network scale. Bigger networks store exponentially more patterns without becoming less reliable at retrieving each one.

The topology does the work. A fully-connected network wastes coupling capacity on redundant connections. The honeycomb gives each oscillator exactly the neighbors it needs to define a phase-locked state, and no more. Structured sparsity creates capacity that density cannot.

The practical test: charge-density-wave oscillators in hardware confirm the theory. This isn't just mathematical possibility — it's physically realizable. Neuromorphic memory at exponential scale, without the linear ceiling that Hopfield hit in 1982.