Sci-Fi Labs · App
Knowledge Graph
Memory as a canvas you can walk. Not a chat log you scroll.
A chat window stores a transcript. This stores a place. Conversations, photos, and voice notes sit as objects on an infinite canvas of time. Scroll and you move through days. Speak and you query or add. The record stays on your machine.
What it is
- Spatial UI. Three.js infinite canvas with time layers. Pan, zoom, scroll through memory.
- Local intelligence. On-device LLM, vision, Whisper, and embeddings on Apple Silicon.
- Living graph. Entities and relationships extracted from conversations, photos, and voice.
- Rich images. EXIF, reverse geocoding, and local vision descriptions become graph nodes.
- Voice-first. Speech enters the same knowledge loop, fully offline.
Chat stack vs this canvas
A chatbot is a lease on a log. You type. It answers. The memory is the thread. Knowledge Graph is the opposite product: a private spatial canvas. You move through it.
| Chat stack | Knowledge Graph | |
|---|---|---|
| Memory | A transcript in someone else's window | Photos, notes, and voice on a canvas you own |
| Retrieval | Top-k from a chat index | Walk the graph. The month. Every photo. Every note. |
| Product | A window you type into | A place you navigate |
What is live on this machine
On this 16 GB laptop the model is Bonsai. Whisper hears. Kokoro JS speaks. Full conversations stay local.
Open WebUI is the temporary surface. Same models. A chatbot window. It was the fast way to get streaming, Whisper, and voice working. Next step is not another engine. Swap that window for this canvas. Same Bonsai. Same Whisper. Same Kokoro JS. The bubbles go. The place stays.
Direction
The long-term surface is cinematic and local. Hierarchical expansion of related knowledge. Time as a navigable dimension. Ambient presence instead of transactional chat. Web and local use now. The same spatial language is meant to extend into XR.
A Sci-Fi app in the strict sense: the interface you were told belonged to movies, running on a machine that never has to phone home.
Want the technical stats: models, architecture, live memory measurements, token throughput, idle versus active load? Field notes: Local AI + RAG: the Full Stack in 16 GB.