Μ Mnemosyne
Sign in

Mnemosyne — how to use the system

1. Install the model

Mnemosyne is free to distribute and runs on CPU (well under 200 MB VRAM; ~6 ms/sample quantized).

git clone <your-internal-mirror> Mnemosyne
cd Mnemosyne
python -m mnemosyne.train          # warm start (core + rf + spectrogram)
python -m mnemosyne.demo          # interactive demo (learn radar + gesture)

2. The free tier

Fresh weights run unlimited for one session (30 min of fused runtime). After that the Enclave's governor arms inside the weights:

  • sensor-fusion latency runs at 2.5×
  • fusion is limited to 2 ports

The governor travels with the model weights (save → distribute → reload, still governed). Re-downloading resets it — that's the research escape hatch. Production deployments carry license ports instead.

3. Seats & subscriptions

Subscribe to seats on the dashboard (billing runs through CAuth + Stripe):

SeatPricePortsSpeed
Development$9/mo41.2× unthrottled
Production$49/mounlimited1.0× full speed

Each subscription allocates seats into your account. Seats are consumed by checking out a license port.

4. Check out a license port

  1. Open the dashboard, allocate a seat.
  2. Press Check out on an available seat.
  3. Set the lifespan (default 30 days, max 365).
  4. Copy the returned .mnport text into a file: <seat>.mnport.

The checkout binds the seat — one port, one seat. The same file (or a copy) is refused by any other installation.

5. Install the port into your model

python -m mnemosyne.enclave_cli install <seat>.mnport --days 30
python -m mnemosyne.enclave_cli status
# unlocked=true  unlimited=true  speed=1.0x  ports=unlimited

The Enclave verifies the issuer signature, the lossy-copy checks (canonical base64 + HMAC + checksum), and the seat binding before unlocking.

6. CLI reference

python -m mnemosyne.enclave_cli issue   --kind dev|prod [--out file]
python -m mnemosyne.enclave_cli install <port> [--days N]
python -m mnemosyne.enclave_cli status
python -m mnemosyne.enclave_cli corrupt <port>   # demo lossy-copy protection
python -m mnemosyne.enclave_cli demo              # full lifecycle demo

7. Architecture notes

  • Backbone: g3-style NodeUnit of VHDUs (causal selective SSM), frozen after warm start.
  • Fuser: g4-style FuserBridge sandbox ladder that adapts any sensor dimension into the latent space.
  • Enclave: session budget + governor in the weights + license-port validation (see mnemosyne/enclave.py).

Honesty note: the Enclave deters casual sharing and raises the cost of production deployment; it is not DRM. Research remains frictionless by design.