LifeOS

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for LifeOS

Inspiration

Most AI assistants forget you the moment a chat ends. To get help, you also send your personal life to someone else's servers, with little say in what is kept or what the assistant does on your behalf.

We wanted the opposite: an assistant that gets more useful over time, but where you can see everything it remembers and approve everything it does.

What it does

LifeOS is a personal AI companion that remembers you, learns your routines, and only takes action with your approval.

• Memory you control. After a chat, LifeOS suggests what it learned. You edit, reject, or approve each item. Nothing is saved without your say. Every memory is a plain text file you can open, change, or delete.
• Everyday tools. It can search your notes, manage a task list, and look things up on the web. Anything that changes data asks for your approval first.
• Reusable routines (skills). Save a repeated task once and run it again with one click. Routines still need your approval for any change.
• Morning brief. A short, read-only summary built from your memories, tasks, and one web search.
• A full activity log that records every action it takes, and what you decided.

How we built it

• Model: NVIDIA Nemotron, called through Nebius Token Factory. [Confirm this line matches your final deployment.]
• App: Python with FastAPI, a simple web page, and a command-line version that share the same core.
• Storage: plain files only (readable text and simple lists), with no database, so users can inspect their own data.
• Safety by design:
  • Approvals and memory confirmations are checked on the server, not just in the web page.
  • File access is limited to each user's own folder.
  • Tool results (like web pages) are treated as untrusted, so the model is told never to follow instructions found inside them.
• Public demo: each visitor gets a private, temporary copy of sample data, with rate limits and a daily cap on model use. The demo is for sample data only.
• Open source: MIT license, with a setup guide written for non-technical readers.

Challenges we ran into

• Making "ask first" a real rule. A button in the page is easy to bypass, so we moved the checks to the server and added tests to prove arbitrary text cannot be saved.
• Clean memories. Early tests saved stray fragments, such as a lone date split away from its event. We tightened what the model may propose so each memory is a single fact the user actually stated.
• A safe public demo. Sharing one demo folder with every visitor would be unsafe and costly, so we added isolated sessions, rate limits, and usage caps.
• Model behavior. Chat answers, tool calls, and memory proposals need different settings, so we turned reasoning off for the strict-format steps.

What we learned

Trust comes from visibility. When people can see what an assistant remembers and did, and can undo it, they use it more confidently. We also learned to treat safety checks as part of the product from day one, not something added at the end.

Honest limitations

LifeOS is an early build. Chat text is sent to the model provider, and this version does not yet hide personal details before sending it. The public demo should be used with sample data only.

What's next

• Hiding sensitive details (emails, phone numbers) before they reach the model
• More tools and calendar support
• A version that runs fully on the user's own computer
• Better skills that suggest themselves from repeated tasks

Try it

• Code (MIT license): https://github.com/regann29/LifeOS
• Video: https://youtu.be/ErGdUE2nIvU

Built With

• ai-agents
• css
• docker
• fastapi
• html
• javascript
• llm
• nebius-token-factory
• nvidia-nemotron
• openai-sdk
• python
• tavily