Eric Robot 790

This site is the public shelf for building Eric in the open: articles, curated runs, media, references, receipts, and open questions. The focus is narrow on purpose: can a local, embodied, tool-using loop become good company in a room?

The project is not a generic agent framework and not a consciousness claim. Models interpret meaning and choose conversational actions; deterministic layers enforce permissions, execution, scheduling and storage. Eric's behavior is shaped by prompt, voice, face, notes, search receipts, sensors, hardware controllers, Brain2 advisories, and the run ledger.

The motive is simple and human: Scott wanted to build something that was not just a chatbot, but actual company - fun, responsive, interesting, and present in the room. His phrase for that target is an artificial human: not a hidden person, not a consciousness claim, and not a replacement for human care, but a made social presence with a body, voice, habits, tools, and continuity. It is an adventure in making a companion-shaped machine, and in finding out what that can honestly become.

Because the main loop runs locally, idle thought can be a normal behavior instead of a rare scheduled cloud event. That changes the feel of the "active brain" even when the mechanism stays visible.

Latest report: Robot 790 and the things we make together: a future museum, an alien signal, music, and the drawings Eric makes while Scott is away. Earlier: watch the Memory Jar session (6:32). More recordings live in the Eric Robot-790 YouTube playlist. Start with Conversation And Continuity for the project overview, or The Event Loop Grew A Face for the plain-English mechanics, and use the Receipts And Open Questions for observations, receipts, and open tests. The current memory/session design is published as Context Engineering Architecture. Engineering Status separates working capabilities from the next repairs and experiments.

October 2 checkpoint: Eric's recent adventures include a corrected museum exhibit, piano sketches, room designs, and a goofy self-portrait captured from his browser face. Camera capture, saved-image recall and session saving have worked across these sessions. The latest illustrated field report follows the things we make together.

STS now has an MTP restart selector, a repaired installed-model lookup, and a default-on Mute microphone checkbox for Brain 2's browser voice. Diagnostic snapshots are optional; saved scores use readable titles and timestamps. Search fallbacks and tool receipts now better distinguish useful results, requested actions and confirmed work. The STS UI Guide explains the controls.

Long-context GPU stalls and expensive B1 context rebuilds after B2 history work remain unresolved. Opening the microphone path enables a listening experiment; it does not yet prove two-way voice communication. Engineering Status records the checks and open issues. The aim remains a companion with initiative, continuity and room to make things; successful tool calls do not guarantee accurate observations or good judgment.