There is a difference between recording something and remembering it. A photograph stores. A brain — a Hopfield network — reconstructs. It takes a broken fragment, a half-corrupted signal, a smear of noise where a face used to be, and it pulls the original pattern back from the field.
Pattern Retrieval is not a collection of images. It is a collection of acts of recall. Each token is a live process: 94 printable ASCII characters stored in a mathematical memory — shattered, corrupted, and then painstakingly reconstructed, over and over, in real time, onchain.
In 1982, physicist John Hopfield published a paper that changed how we think about memory. He described a network of artificial neurons — binary units, firing or silent — where memory was not a file stored in an address but a valley in an energy landscape.
Imagine terrain. Rolling hills, deep basins. Each stored memory carves a basin into this terrain. When you show the network a corrupted input — a pattern broken by noise — the network's dynamics are like a ball placed somewhere on the terrain. It rolls downhill. It settles in the nearest basin. It finds the closest stored memory and locks onto it.
That is retrieval. Not lookup. Not search. Settling.
The weights wᵢⱼ between neurons encode the memories via Hebbian learning — neurons that fire together, wire together. The state vector s is the current pattern. The energy E descends with each update step. When it bottoms out, you have your answer.
The network doesn't know it's recalling. It simply moves toward stability.
Every piece in Pattern Retrieval begins as noise. Not stylized noise. Not aesthetic noise. Real, catastrophic, information-destroying noise — bits flipped at random, the original character's pixel structure scrambled beyond recognition. This is not metaphor. This is the actual computational process running in your browser.
What you watch onscreen is a Hebbian recall loop: the network repeatedly applying its weight matrix to the corrupted state, nudging each neuron toward the configuration its learned memories prefer, reducing energy step by step. Each pass, the character becomes slightly more itself. At pass 8, it arrives.
The artist chose λ = 0.35 — a temperature parameter governing how aggressively the network updates. Too high, the network overshoots. Too low, it stalls. At 0.35 it walks the edge: recall that looks like struggle. Which is exactly what memory feels like from the inside.
Most generative art uses randomness as decoration. Pattern Retrieval uses it as the problem to be solved. The noise isn't texture sprinkled on top — it is the starting state of the system, and the artwork is the process of escaping it. What you are watching is computation recovering meaning from entropy.
Printable ASCII runs from character 32 (space) to character 126 (~). 94 printable, non-space characters. Every one of them is in this collection. Not as static images — as live recall events, running continuously, each on its own loop, each carrying its own phosphor color and harmonic audio signature.
The characters are sorted into six families, grouped by their visual structure — the geometry their pixel matrices share. The taxonomy is not arbitrary. It reflects how the network experiences them: characters with similar pixel distributions occupy nearby regions of the energy landscape and risk confusing each other during retrieval. Dense Symbols — complex, pixel-heavy glyphs — were hardest to train. They received a distinct targeting-display aesthetic as recognition of their difficulty.
Most NFTs are pointers. They are tokens on a blockchain that point to a JPEG hosted on a server somewhere. If that server goes offline, if that company folds, if the IPFS node unpins — the artwork is gone. The token remains. The art does not.
Pattern Retrieval is stored in its entirety on the Ethereum blockchain. The code that generates the recall animation, the weight matrix encoding all 94 characters, the audio synthesis engine, the CRT shader — everything lives on-chain. The Ethereum network is the server. As long as Ethereum exists, these pieces run.
This is not a technical detail. It is the conceptual core of the work. A piece about memory — about information surviving corruption and noise — must itself be corruption-resistant. The medium is the message. The permanence of the chain is the permanence of the recalled pattern.