The model you talk to today is the same one that was shipped months ago. It can't learn from you. Whatever you explained to it yesterday is gone this morning, and the only way to genuinely teach it something new is a full retraining run in a data centre you don't control. Worse, that retraining run tends to overwrite what the model already knew. The industry has a name for this: catastrophic forgetting.

We think this is the single biggest limitation of today's AI, and it is a limitation of the architecture, not of scale. Bigger frozen models are still frozen. So we started again from the neuron up. The result is Voltaic: a brain-inspired network that learns continuously while it runs, protects what matters, and can show you exactly which knowledge is safe.

Why frozen models forget

A conventional neural network stores everything it knows in one shared set of dense weights. When you train it on something new, gradient descent nudges those weights towards the new task without any notion of which ones were holding up the old tasks. Learn French, lose some German. Learn a new product catalogue, start hallucinating the old one. The workarounds are expensive and partial: replay buffers that keep re-feeding old data, frozen layers that stop learning altogether, or simply retraining from scratch on everything every time.

Brains don't work like this. You can learn to drive a new car without forgetting how to ride a bike. Biology solved continual learning with sparse, event-driven computation and with mechanisms that decide which connections to harden and which to leave plastic. Voltaic borrows both ideas.

Spikes, not dense activations

Instead of dense, continuous activations, Voltaic processes information as discrete spike events, the same sparse, event-driven dynamics your neurons use. It is built from binary spikes, leaky integrate-and-fire (LIF) and Adaptive-LIF neurons, and STDP-weighted attention, running across 16 simulated timesteps.

  • Binary spikes. Every input, whether text, an image, audio or structured data, is encoded into binary spike tensors over simulated neural time. A neuron either fires or it doesn't.
  • Leaky integrate-and-fire neurons. Each neuron accumulates incoming signal, leaks a little over time, and fires only when it crosses a threshold. Most neurons are silent most of the time, which is why spiking networks are so efficient: they only do work when there's something new to process.
  • STDP attention. Spike-timing-dependent plasticity strengthens connections between neurons that fire together in the right order. Voltaic uses it to weight attention inside a spiking transformer, so the model's notion of what is relevant is shaped by timing, the way biological attention is.

Three things fall out of that design

  • Learn live. Teach facts, skills and gestures at inference time. There's no retraining run and no redeploy; knowledge lands while the model is running. Tell it about a new customer in the morning and it can use that in the afternoon.
  • One brain. Text, voice and camera flow into the same weights. It is not three APIs stitched together but a single spiking representation for every sense. A gesture taught through the camera and a fact taught in chat live in the same mind.
  • Prove it. Hierarchical Synaptic Consolidation hardens important synapses. We can show you the forgetting curve, which knowledge is protected, and why. Continual learning you can't inspect is just drift; Voltaic's is measurable.

The loop

Every interaction goes through the same four stages:

  1. Encode. The input becomes binary spike tensors over simulated neural time.
  2. Spike. LIF and Adaptive-LIF neurons fire across 16 timesteps through the spiking transformer, with STDP-weighted attention.
  3. Consolidate. Living Memory, driven by HSC v2, tracks each synapse's importance, velocity and clarity, so persistence has to be earned rather than granted to everything.
  4. Answer. ContextFlow assembles the relevant memory. The model answers, explains itself and keeps learning from the exchange.

Because consolidation happens on every pass, learning isn't a separate phase that happens somewhere else. It's part of using the model.

What the numbers say

We test these claims on public continual-learning benchmarks against strong dense baselines, without a replay buffer. On CORe50-NI, adding HULL brought forgetting down from 1.20% for a standard model to 0.04%, while average accuracy rose from 95.01% to 95.23%. On Split ImageNet-1k, Voltaic retained 4.2× more knowledge than dense 7B baselines, and on Split CIFAR-100, 2.6× more. Across tasks, the average drift is 0.71×, which means old skills often sharpen as new ones are learned. The full protocol is in our benchmarks write-up, and every claim is backed by 647 automated tests, all passing.

Small enough to own

Voltaic runs at 355M to 7B parameters on a single GPU today. That matters as much as the accuracy figures: a model that learns from your data should run where your data lives. And because spiking neurons only fire when there's something new, the architecture is built to run on Verkko's own proprietary hardware, which is where we're heading next.

When a task genuinely needs frontier scale, Voltaic doesn't have to answer alone. Its core can route to an open model on your own hardware or a frontier model in the cloud, weighing confidence, context sensitivity and complexity, and fuse the answer with its own memory. The spiking core stays in charge of what is shared and when.

Embodied Living Intelligence

This is what we call Embodied Living Intelligence: a cognitive engine built to live inside a body, not just a chat window. It remembers what it did yesterday, adapts to what's different today, and evolves without a retraining run. Today it powers Verkko Nexus for organisations and Verkko Halo for people. Next, through Verkko Robotics and Verkko Kinetic, it gets a body: robots that learn skills on the job without unlearning how to walk.

You can try Voltaic today at voltaic.live, or read more on the Voltaic page.