Charlot Lab · The Adaptive Unit
A Transformer's memory is its context window — finite, and opaque. The Dragon Hatchling (BDH) does it the brain's way instead: its working memory lives in synapses that change while it runs, strengthened by a Hebbian rule whenever it reasons about something. Learned during inference, with no training and no gradient — and you can point at the exact synapse that holds a fact. Here is that mechanism, on your device.
the neuron particles · a synapse brightens as it learns
1 · teach it a fact (fires a Hebbian update, live)
2 · ask it — activate a concept, let it spread through the synapses
—
the synaptic memory · every cell is one readable synapse
Teach it a fact and watch one synapse brighten. Then ask: activate a concept and the signal spreads down the strong synapses to what it was paired with — recalled from a memory that was written while it ran, not trained in. Because the memory is a synapse, not a hidden vector, you can read it straight off the grid — which is why BDH is monosemantic and interpretable by construction. Teach a new fact and it simply adds another synapse: unbounded memory, no forgetting, no context window.