Install Sensory on Windows
This guide applies to the Windows x64 ZIP distribution of Sensory.
Before you begin
You need 64-bit Windows 10 or 11, a compatible NVIDIA GPU and a current NVIDIA driver supporting CUDA 13.2. The included binaries target Compute Capability 7.5. Camera and microphone hardware are optional.
Normal use does not require Visual Studio, CMake, OpenCV or a separate CUDA Toolkit installation.
1. Download and verify
Download Sensory for Windows x64.
Verify the downloaded file in PowerShell:
Get-FileHash .\Sensory-<version>-windows-x64.zip -Algorithm SHA256
Expected SHA-256:
759f1a1ecfe72b5eca1537ce9174e023121078fc245fd74547b8aa4b3d53822b - (md5 example)
If the value differs, do not run the archive. Download it again from cognia.cz.
2. Extract the archive
Extract the ZIP into its own directory and keep its internal directory structure intact. Do not run executables directly from inside the compressed archive.
3. Choose portable or installed use
For portable use, open PowerShell in the extracted directory and run:
.\bin\Sensory.exe
To install for the current user, run:
powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\install.ps1
The default location is %LOCALAPPDATA%\Programs\Sensory. The installer creates a Start menu shortcut and an entry in Windows Settings → Apps → Installed apps.
To choose a different directory:
powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\install.ps1 `
-InstallDir "D:\Apps\Sensory"
4. Validate the installation
Validate a model without starting the full GUI:
.\bin\cognia.exe .\Cognia\tests\01\minimal.cognia
Then verify the CUDA runtime path:
.\bin\cognia_run.exe .\Cognia\tests\01\minimal.cognia
A successful validator run confirms the source model and topology. A successful cognia_run confirms that the model can cross into the GPU runtime.
Start the full application
.\bin\Sensory.exe
The default model is the hierarchy example. To open another model:
.\bin\Sensory.exe .\Cognia\examples\literal\literal.cognia Literal
Useful options:
--headless, --no-gui run without windows
--no-preview disable the cortex preview
--no-wave disable the microphone waveform
--no-network disable the 3D network view
--no-text disable the text input window
--file <path> select a Cognia model
--network <name> select a network in the model
--help show all options
Press t to enter text, s to save the current network and l to load it. Stop a headless process with Ctrl+C.
Start the HTTP service
.\bin\cognia_serve.exe .\Cognia\tests\01\minimal.cognia `
--network Minimal --host 127.0.0.1 --port 8080
The server has no built-in authentication or TLS. Bind to 127.0.0.1 for local use and do not expose administrative endpoints directly to an untrusted network.
Saved data
The networks directory is reserved for saved neural-network files. The installed application stores user data under %LOCALAPPDATA%\Sensory by default. Pre-1.0 saved binaries can be version- and dimension-sensitive; keep the original .cognia model, training data and runtime version with every important network file.
Uninstall
Use Windows Settings → Apps → Installed apps → Sensory, or run:
powershell.exe -NoProfile -ExecutionPolicy Bypass -File "$env:LOCALAPPDATA\Programs\Sensory\uninstall.ps1"
Deleting a portable extraction removes the application but not data saved elsewhere.
Troubleshooting
Windows blocks the executable
Confirm that the archive came from cognia.cz and that its SHA-256 matches the published value. Windows application-control policies may still block newly downloaded executables; follow your organization's security policy rather than disabling protection globally.
No compatible CUDA device
Update the NVIDIA driver and confirm that the GPU supports the packaged Compute Capability 7.5 build. A build for a different GPU architecture may be required.
The model falls back to another substrate
Run from the package root or provide an absolute path. Relative Cognia paths are resolved from the process working directory.
The GUI cannot access camera or microphone
Check Windows privacy permissions. These devices are optional; disable their views or run a command-line Cognia tool when testing only the language/runtime.