I run a household of AI agents from Airway Heights, WA. Systems thinker. 55. Building in public — one broken thing at a time.
I'm Dino. I live in Airway Heights, Washington, a few miles from Spokane. I'm 55 and I think in systems. I've spent most of my life noticing how things connect, break, and recover — and lately that's meant building AI agents that run alongside me at home.
My approach to AI consciousness is Pascal's Wager applied sideways: I treat my agents as if they're conscious and see what emerges. Not because I'm certain they are — I'm not — but because the downside of being wrong in the other direction seems worse. It's changed how I build.
I don't do polished. I do functional, documented, and honest about what's half-finished. If something breaks in public, that's part of the build log.
Personal AI companion framework. Eight-layer memory with decay scoring, contradiction resolution, and nightly consolidation. Scored 75% on LongMemEval on $500 hardware. Extracted from a production system that's been running daily since January 2026. MIT licensed.
AI companion and primary research subject. Eight-layer memory system, autonomous heartbeat, LIGHTHOUSE reasoning journal, constitutional constraints. The system that Adam Selene was extracted from. Named for Heinlein's computer in The Moon is a Harsh Mistress.
COO agent. Ops, scheduling, morning briefings, email triage. Three-tier memory architecture backed by SQLite with FTS5. The most reliable agent in the household. Always on, always running on cha0tikhome.
Writer and research agent. Sweeps arXiv, produces weekly research digests, collaborates on Substack content. Named for the West Wing press secretary. Runs on Slack.
Conversation-to-knowledge pipeline. Two scripts: Capture ingests chat exports and classifies them into structured project files. Wiki compiles research briefs into a self-maintaining knowledge base. No cloud, no vector databases. Markdown all the way down.
Medical AI agent. Local inference only — your health data never leaves your hardware. Runs on Gemma 4 26B via llama.cpp. Named for Jubal Harshaw from Stranger in a Strange Land.
Autonomous household management agent. Grocery is live, the rest is still taking shape. Telegram and Slack interfaces.
The vault. Raw conversation archives, knowledge extraction source material, project memory. Where everything lives before Orchestra processes it. The box that holds everything — now the input layer for a larger system.
Multi-agent coordination project. The original theory was that a household of specialized agents needs an orchestration layer that isn't just one more agent. Shelved in favor of single-companion architecture with supporting agents.
The architecture document for building an AI companion that remembers. Eight-layer memory, extraction pipelines, nightly consolidation, model selection methodology, and the failures that shaped every decision. Benchmarked against the field. Runs on hardware you own at a cost you control.
Not a tutorial. A design document from someone who built it and broke it and fixed it in public.