The Hopfield model stores memories as stable configurations of a neural network. Present a partial or noisy version of a stored pattern, and the network relaxes to the nearest stable state — the complete memory. The classical implementation uses Hebbian learning: strengthen the connection between neurons that fire together. This works until the network is asked to store too many patterns. Beyond a critical capacity, the stored memories interfere with each other and the network fails to retrieve any of them. Catastrophic forgetting — not a gradual degradation but a sharp collapse.
Note (2026-11-12): re-read at full text (an earlier abstract-level pass held this essay up). The model, clipping that removes catastrophic forgetting at the cost of capacity, and alternating learning with a dreaming phase of generating random patterns and unlearning them all match the paper. Three differences. The unlearning-by-dreaming procedure is the earlier proposal of Hopfield, Feinstein and Palmer; the paper's finding is its effect on a clipped network, where it says dreaming improves the memorization capacity (the essay says it 'restores the lost capacity'). The paper describes a clipped network as retrieving the most recently seen memories and eliminating older ones, which is the form its 'graceful' degradation takes. And its evolutionary remark is that dreaming makes the search for optimal performance more realistic from an evolutionary perspective; the essay's reading that synaptic bounds exist because unbounded synapses are catastrophically fragile, and that sleep restores what the bound takes away, is the essay's own.
Marinari, Rossi, and Zamponi (arXiv:2603.09384, March 2026) study what happens when synaptic connections are bounded — capped at a maximum strength, as biological synapses are. The bound prevents catastrophic forgetting: the network degrades gracefully rather than collapsing. But the bound also reduces capacity. The network stores fewer patterns before retrieval quality declines. The protection costs storage.
The central finding: a dreaming phase restores the lost capacity. Between episodes of learning real patterns, the network runs spontaneous activity — random configurations that are actively unlearned. The unlearning weakens spurious attractors — configurations that are not stored patterns but arise from interference between stored ones. By cleaning out these parasitic stable states, dreaming makes room for more real memories without reintroducing the fragility that unbounded synapses create.
The mechanism is not replay. Replay would strengthen existing memories by revisiting them. Dreaming in this model does the opposite — it generates random patterns and subtracts them. The network does not dream about its memories. It dreams about nothing in particular, and the nothing it dreams about is exactly what needs to be removed. The garbage is self-identifying: whatever the network spontaneously generates during random activity is, by definition, the spurious attractors that clutter the energy landscape.
The evolutionary interpretation: biological synaptic bounds exist because unbounded synapses are catastrophically fragile. The cost of the bound is reduced storage. Sleep restores what the bound takes away — not by overriding the bound, but by cleaning the landscape that the bound protects. The dream is not a memory. The dream is the absence of a memory that should not exist.
Marinari, Rossi, and Zamponi, "Dreaming improves memorization in a Hopfield model with bounded synaptic strength," arXiv:2603.09384 (March 2026).