Skip to main content
ngram — white globe mark on periwinkle Give an agent continuity beyond a conversation. Build its memory, connect its tools, and bring it into the room. ngram is an open engine for persistent agents. The Entity holds identity, memory, relationships, and judgment. The harness runs it. A shell gives it a body. You can change the model or the surface while keeping the same Entity.

Choose your first build

  • Quickstart: choose hosted or local inference, create an identity, and start a conversation.
  • Spatial setup: connect a WebXR shell and build a shared scene.

Continuity is the architecture

A message from Telegram and a request from a headset reach the same running Entity. Its inference provider generates responses; its runtime owns the state that makes those responses part of a continuing history.

Shape an identity

Define traits, voice, knowledge, and the way your Entity meets people.

Build in the room

Combine spatial tools into diagrams, demonstrations, and shared environments.

Understand the runtime

Follow a turn through cognition, retrieval, tools, memory, and presence.

Operate deliberately

Control background activity, context, cost, and when an Entity stops.

Start small. Keep the same Entity.

You do not need a headset, GPU, Railway deployment, or Cloudflare tunnel to begin. A hosted setup runs the Entity on your computer and sends inference requests to your chosen provider. A local setup runs those requests through Ollama. Add messaging, spatial embodiment, or an always-on deployment when your project needs it. The core concepts explain what moves between those environments and what stays with the Entity.
These docs describe the shipped v1 runtime. Runtime features and registered tools depend on configuration and the connected surface. Check surface capabilities before assuming headset tracking or a backend is available.