Sensor fusion that learns
new senses it has never seen.
Mnemosyne is a small recurrent fusion core that warm-starts on a few known sensors — then learns a brand-new modality in seconds by adapting only a small port, with the core and fuser frozen. Weights are free to distribute; license ports unlock the Enclave for production.
sandbox ladder
How it works
Warm start
The core + fuser are trained on two known sensors (RF ranging, audio spectrograms) so the fusion pathway already knows how to integrate a new voice.
Learn a new sense
Hand it radar or gesture data it has never seen. A fresh random port adapts in ~40 s with the core frozen — 13× fewer parameters than training from scratch.
Deploy, licensed
Weights ship free. The Enclave grants one unlimited session, then a governor slows fusion and caps the port count — until a license port from this site unlocks full speed.
Check out ports
Allocate dev or production seats from your subscription, then
check out .mnport license files. Install one into your
model and the Enclave unlocks it for the seat's lifespan.
The demo arc
python -m mnemosyne.demo
PHASE 1 — KNOWN SENSORS ......... rf + spectrogram: acc 1.000
PHASE 2 — NOVEL SENSOR: RADAR ... solo acc 0.23 → 0.84 (40s, core frozen)
PHASE 3 — FORGET & RELEARN ...... gesture solo 0.14 → 0.94 (35s)
PHASE 4 — PERSISTENCE ........... fusion.gguf, verified by a C loader
PHASE 5 — ENCLAVE ............... free session → governor → license unlock
python -m mnemosyne.demo --enclave # the full licensing lifecycleSeat plans
- 4 fused ports
- 1.2× (unthrottled) speed
- 30-day license ports
- For researchers & developers
- Unlimited fused ports
- 1.0× full speed
- 30-day license ports
- For production deployments
Subscriptions are handled through CAuth + Stripe. Allocate seats from your tier allowance, or buy single seats.