The experimental question
Eight cells form a periodic ring. Observations arrive at specified cells: a known key symbol and a fresh eight-bit value. Later, a query starts at a chosen origin and must recover the value by travelling through neighbouring cells.
During evaluation, the shared neural parameters are frozen. The only acquired information is written into private per-cell memory. The test asks about learning new associations inside that state—not learning new language concepts or updating a general-purpose model.
One rule, two kinds of stored information
The supplied fast-memory checkpoint has 385 shared parameters: learned key vectors, a value encoder, a decoder and an output scale. Each cell carries an additional 8 × 16 matrix, or 128 adaptive floats. That matrix belongs to the current episode rather than the shared checkpoint.
Keys are drawn from 16 known symbols. Values and their assignments vary between episodes. The shared representation is learned during outer training, while the local write law and communication protocol are deliberately supplied.
A local write corrects a prediction
The model normalises a key vector k and encodes the observed value as v. At the writer cell, the matrix A predicts A k. A normalised delta update corrects that prediction:
Only the observation site changes. This is a structured fast-weight update, not a newly discovered learning law. Correlated key vectors can interfere with one another, which is why key width and learned geometry matter.
A query makes one lap
A fresh packet contains a key and an initially empty value accumulator. In each synchronous update, it moves one cell forward and reads that cell’s contribution. After eight updates, the packet returns to its origin and the decoder produces eight output bits.
No expected answer is supplied to the query function. The browser keeps observed labels only for its comparison display. Queries are read-only: they do not alter private memory or the frozen shared weights.
Intervene, rather than just watch
The browser lab lets you write and overwrite values, step a query around the ring, erase a writer’s matrix, erase all private memory, or disable neighbour relay. These controls make the source of an answer visible.
For a clean relay intervention, choose a query origin different from the target’s writer. With relay disabled, the source protocol repeatedly reads the origin instead of receiving remote contributions. The browser is a hands-on mechanism demonstration, not a new statistical benchmark.
Repeated writes to one key must stay at its original writer cell within a session. Moving an association elsewhere without removing its previous contribution is outside the overwrite protocol; the interface rejects that operation rather than silently creating a duplicate.
What the recorded result establishes
The learned 16-dimensional-key arm achieved 100% whole-value exact recall on the recorded eight-record evaluation, averaged across three initialisation seeds. Erasing memory or disabling communication reduced the controlled task to approximately chance. The results page provides the full conditions and comparison arms.
An exact symbolic dictionary also solves this task. The controls are not matched in parameter count, compute or tuning budget, so these results should not be read as a general ranking of neural architectures.