Skip to main content
The recommended setup gives your ngram hosted thinking and embeddings, a Railway worker, a dedicated Postgres memory database, and a persistent Linux workspace. Your browser or headset connects to that same Entity.

​
The guided path

  1. Create an Entity or choose an existing configuration.
  2. Choose a provider and enter its API key. OpenAI includes a chat default and text-embedding-3-small for 768-dimensional memory. The wizard checks the model catalog and sends a small embedding probe.
  3. Link a Railway project. Review the named worker, memory database, and /app/data volume before deploying. The wizard creates missing resources and sends credentials privately through the CLI.
  4. Pair an authenticated worker endpoint and verify the connection. Open the AR app with the exact command printed by setup.
Railway hosting and provider API usage are billed by those services. You need Python, uv, Node.js 22+, a provider account, and a Railway account. A local GPU and Cloudflare are not part of this path.

​
In the application

Choose New ngram to open the guided experience. Pair a saved worker or enter its URL and pairing token, choose a body and voice, then connect. The app verifies the authenticated worker before saving a body. A failed connection leaves your form open for correction.
The New ngram wizard in light mode, with a periwinkle illustration on the left and the recommended cloud setup on the right. Its four steps are Home, Connection, Presence, and Ready.

New ngram starts with the recommended cloud home: a provider API for thinking and embeddings, a Railway worker for memory and files, and your browser or headset for spatial presence.

Cloud setup saves pairing details automatically. If you ran setup while the app was open, use Refresh connections. Credentials stay in ignored server runtime files; shell YAML contains only an opaque connection reference. A new body on the same worker is the same Entity, with its existing identity and relationships. For a separate ngram, create a separate worker and memory database through cloud setup.

​
Memory is part of the default

Embeddings power semantic recall. Postgres holds persistent experience and relationship data. The worker volume keeps knowledge, journal, body state, and files available across redeployments. Keep both storage resources attached. Pairing an existing worker preserves its memory settings. An OpenAI chat connection with existing embeddings can still depend on a home gateway. Equal vector dimensions do not make two embedding models compatible. The app blocks an onboarding switch to a different embedding model when existing memories are detected; migrate those memories deliberately before changing their vector space.

​
Other profiles

Use uv run ngram setup --profile hosted, --profile local, or --profile hybrid when you explicitly want these alternatives. Cloud setup does not overwrite a deployed Entity’s inference or memory configuration; use uv run ngram ar setup to pair an existing worker.