Live and interactive in the browser: send a change and watch the network relearn in real time. Each one runs the actual learning core, not a recording.
Evidence
Measured results, not projections
Bounded across thousands of distinct networks and billions of live-learning steps, with zero divergences.
What’s demonstrated
One kernel, many tasks
The same kernel, never edited per task, was dropped into five different control problems: lane keeping, unstable balance, foraging, sparse sequential goals, and self-taught landing. It was bounded in every one, whether the teacher was a coach, a reward, or the world grading its own prediction. The update and envelope remain the same.
Survives what tuned controllers don’t
Flip the steering polarity mid-drive and classical controllers flatline. The learner re-adapts and stays in envelope. Scramble its internal state and behavior comes back.
Deploy blank OR trained
No reward function, no training data. A blank network babbles, teaches itself the plant from its own sensor stream, and flies it, with the goal as a runtime query that never enters the weights.
Refuses to learn a dying machine
On 260 engines from NASA’s C-MAPSS turbofan benchmark (data only, not a physical machine), the consolidation gate keeps tracking benign drift and refuses to fold degradation in as the new normal once it begins.
Learning on real silicon
The kernel runs on a five-dollar microcontroller with weight updates happening on the chip: a full sense, act, and learn cycle at tick rate, in tens of kilobytes of RAM, in plain C++ with no framework or OS.
Grows without damage
A deployed network can add neurons mid-stream and the new capacity trains up as if it had always been there, with held behavior intact and the bound holding through every graft.
Learns from irregular time
Event-driven sensors wake on their own schedule and the gap between samples is often the signal. The kernel learns directly from that stream and discriminates intervals a time-blind learner has no access to at all.
No floating-point unit
The structural bound sizes a fixed-point format that provably cannot overflow, so the complete learning step, not just the forward pass, runs in 16-bit integer arithmetic.
The comparison
A properly stable network stays bounded and still sheds old skills as it learns new ones. Stability alone doesn’t guarantee retention.
On a shared multi-task stream at the same size, that separation is worth reproducing yourself.
Demos
Route in
Licensing & pilot inquiries
hello@synapseruntime.ai
Don’t just take our word for it. Ask about a pilot and run it on your own hardware.