Releases: ngramspatial/ngram
Release list
ngram v1.0.1 — The Figment Update
Build something you can reach for.
ngram v1.0.1 introduces Figments: creations that carry their own behavior, physical properties, and ways to interact. An agent can shape a model in Blender, bring it into your space, and turn it into a lantern you operate, a mechanism you tune, or a sculpture that reacts when you touch it.
The conversation becomes something you can keep.
Give objects behavior
Figments layer named parts, anchors, grips, properties, and actions onto Blender models or Spatial assemblies. Agents write scoped JavaScript to decide what happens when an object is grabbed, activated, moved, or struck. Five dedicated tools make authoring, physics, behavior, interaction, and publishing part of the existing Spatial suite—now 28 tools.
Make them feel physical
Adjust mass, gravity, friction, bounce, damping, and movement constraints. Build compound collision shapes, sensors, hinges, springs, and motorized mechanisms. Collision events carry measured contact data that programs can respond to.
One- and two-hand grips give creations deliberate places to hold. The native Objects drawer lets people change properties, tune physics, edit behavior, and pause or resume it. Human grabs retain control of an object's pose.
Keep creating in Blender
Blender runs on the agent's execution host and publishes successive geometry previews into Spatial. Your placement, Figment definition, and authored grips survive model revisions. Editable .blend checkpoints stay part of the workflow.
Make it yours. Pass it on.
Publish an immutable local version, export a self-contained .figment.json file, and place independent editable copies. Packages carry assets, code, physics, controls, grips, and optional Blender source. Sharing currently uses files and a local library.
Behavior and simulation run locally, without per-frame model calls. Imported and restored programs start paused; Stop freezes programs and cancels pending package jobs.
Also in this release
- Working named movement and faster walking.
- Improved speech captions, voice settings, and Cartesia support.
- Highlighted slash commands and working YouTube window controls.
- Persistent inference pause, repaired compaction, and support for long-running agent work.
- An antialiased ground grid that adapts to its background.
- MIT licensing across the project.
Upgrade the Entity worker and Spatial client together. Blender belongs on the configured execution host. Back up existing worlds and Blender artifacts; Figment exports are the portable backup for packaged creations.
Validated with 727 Python tests, 128 AR tests, and browser checks for behavior, Stop, publishing, copying, export, and restoration. XR grip paths have automated coverage; headset feel and scanned-room collision remain unverified.
Figments · Physics · Blender · Full changelog
ngram v1 — The same agent, everywhere.
The same agent, everywhere.
An agent should be able to leave the chat window without leaving its history behind.
ngram v1 brings memory, identity, relationships, tools, and spatial embodiment into one open engine for persistent agents. Start a conversation in Telegram. Return to it in a terminal. Bring the same Entity into a spatial environment through WebXR. Across the surfaces you connect, its stored memories and relationships stay with it.
What ships in v1
- Continuity across sessions. Episodic and semantic memory, journals, knowledge, and relationship state give each Entity a history to draw on. Portable
.ngramcontainers let you inspect, export, and move that state. - One identity, many surfaces. Telegram, Discord, the CLI, HTTP, and the browser can connect to the same persistent Entity.
- A body you can change. Shareable spatial shells combine models, animation, voice, and behavior. Desktop and supported WebXR devices provide panels, 3D objects, physics toys, annotations, and media.
- Your choice of inference. Run locally with Ollama or connect a hosted provider. Keep the Entity's identity and memory while changing the model behind it, with local and hybrid deployment paths.
- Autonomy with controls. Configure background activity and tool access, interrupt an active response, and see when context is being compacted.
Start here
With Python 3.11+ and uv installed:
git clone --branch v1 https://github.com/ngramspatial/ngram.git
cd ngram
uv sync --extra dev
uv run ngram setupThe guided setup walks through inference and Entity configuration. Add Node.js 22+ when you're ready to give the Entity a body.
Explore ngram · Read the quick start · Build with us
Welcome to ngram.