GuiOS, a second brain I switched off.
A multi-agent system over a local database of my own life: money, health, projects, friends, reachable from a Telegram bot. It worked. I shut it down in September 2026 anyway.
Why I turned it off
The memory pipeline broke in August when an API credit ran out. I noticed a month later. That was the whole verdict: a system you don't miss for a month is storage, not a brain.
I like keeping my data organized. It turns out that is a different need from wanting to query it. GuiOS answered questions I never asked, and the one place where I did want help, my social life, depended on data it didn't have: the invites that live in group chats and never reach a calendar.
What I kept: the habit of writing decisions down, a deterministic gate before anything the model writes, and a strong suspicion of any personal tool that needs daily feeding.
How it worked
One orchestrator, @soul on Gemini 2.5 Pro, routed each request to a domain agent and kept the daily memory log. It was the only agent allowed to write memory. The domain agents ran on Gemini Flash and could only read the database through a read-only query_db(sql) tool.
@wellbeing
Sleep, training and nutrition from Xiaomi and Samsung wearable exports. Sub-agents for training and health.
@money
Allocation and retirement modeling: Monte Carlo with 5,000 trajectories over 240 months, deflated by IPCA.
@intelligence
Projects, books and courses. Mapped knowledge gaps from years of my own code.
@entrepreneurship
Read meeting transcripts, WhatsApp threads and GitHub activity looking for business signals.
@social
Suggested outings with dates and costs from Google Maps history and chat frequency, with a travel sub-agent pulling live flights and hotels.
Memory
Every conversation, with Claude Code, the bot itself, ChatGPT or Gemini, went in raw and was compressed up a pyramid. No RAG over my own data: agents queried SQLite directly.
Daily
Three sections a day: coding sessions, bot chat, other AI conversations.
Weekly
Closed on Sunday night. What happened, what shifted, what to carry forward.
Monthly
Themes, decisions and recurring patterns.
Quarterly
Durable profiles: values, around 90 relationships, finances, projects.
Twenty-odd tables across finance, health, time use, learning and communications. External data came from public APIs, wearable exports and a flight and hotel scraper.