Read this before comparing numbers

These are the repository’s recorded CPU reference experiments, not measurements taken from visitors’ browsers. The memory and language results belong to separate models. Browser operation timings measure a different implementation and are not directly comparable to the original runs.

The work is exploratory. There was no broad hyperparameter search or matched architectural competition. The full source report retains the controls, temperature samples and training context.

Remote associative memory

Each main arm received 600 outer training updates at batch size 64. Evaluation used 512 fixed new episodes per condition, shared across three initialisation seeds. Scores below are whole eight-bit exact recall, not per-bit accuracy.

Recorded exact recall (%)
Model8 recordsOverwrite16 records
Learned 16D keys100.00100.00100.00
Frozen random 16D keys65.8269.9230.53
128-state GRU control4.6929.172.02
Learned 8D keys55.7962.8920.12
Exact symbolic dictionary100.00100.00100.00

The dictionary is an untrained positive algorithmic control with exact addressing. The fast model has 385 shared parameters; the GRU control has 69,313. They share a 128-float private payload per cell but not a compute cost or a matched training recipe.

Where does the answer come from?

Learned 16D-key memory interventions
ConditionPer-bit accuracyExact recall
Intact memory100.00%100.00%
Erase all private memory49.76%0.20%
Disable writes49.76%0.20%
Erase writer cell50.94%0.39%
Disable neighbour relay49.74%0.39%

Random-value chance is 50% per bit and 1/256 (about 0.39%) for an entire value. Finite sampled results fluctuate. The reset output is deterministically all zero rather than a random guess.

A small language baseline

The language checkpoint has a 17-byte maximum dependency window and was trained on template text. It is not coupled to the fast-memory model.

Sampled validation, original single-seed language runs
Training updatesBits per byteProcessed byte targets
4001.7506614,400
2,0000.46443,072,000

The longer run restarted with the same seed and data schedule, rather than resuming the earlier optimiser state. The evaluation sampled 16 fixed validation batches. There is no tuned language-model baseline in this package.

More iterations did not always help

Applying the same eight-iteration-trained language rule for 16 iterations increased sampled validation loss to 0.6635 bits per byte; at 32 iterations it rose to 1.6321. The theoretical receptive field grew, but quality deteriorated in this checkpoint.

This is an inference-depth intervention, not a comparison between separately trained models. Increasing the number of cellular updates is a hypothesis to test, not a guarantee of better reasoning.

What is not established

These experiments do not establish open-ended language understanding, general continual learning, superiority to transformers, learned routing, human-like memory or production readiness. They do provide small, executable mechanisms with explicit controls and checkpoints that can be inspected and reproduced.

SOURCES & REPRODUCIBILITY

01 · Recorded results, conditions and limitations ↗02 · CellularMemory: source implementation ↗03 · CausalNCALM: source implementation ↗04 · Research guide and architectural discussion ↗
Run the experiments ↗Research programme →