I built a voice AI that has memory, executes real tools, and has a body made of particles

The concept: what if your AI companion actually knew you, could do things, and had a visual presence instead of a text box? Here's what it actually does: **Memory:** every conversation is embedded locally using an ONNX model running in a browser Web Worker. Semantic search surfaces relevant context from past sessions. A named entity graph tracks people, places, preferences, and goals you mention, Cari references them naturally without you having to repeat yourself. **Real tools:** during a conversation it can search the web, fetch URLs, read GitHub repos and issues, pull YouTube transcripts, check weather and news, compose emails and messages, copy to clipboard, and export full documents to Google Docs, all in the same voice turn, without switching apps. **Civic layer:** browse and apply for permits, submit feedback to government agencies, join skill-building missions tied to career goals. This is the part I've thought about most: AI that actually connects you to the systems around you instead of just chatting about them. **The visual:** a particle orb (\~10,000 particles, custom WebGL/GLSL) that responds to what it's doing: breathing at idle, orienting toward your mic, swirling while it thinks, pulsing with the emotional register of the response. When it describes something physical it morphs into a 3D mesh of it. The shape isn't decoration, it's the AI showing its work.