As anyone who has been in a long-term relationship knows, it takes a lot of communication, planning, and logistics to keep a household and two lives running smoothly. At work these days, the idea of starting new projects or maintaining existing work without Claude or other agents seems ridiculous given how helpful this technology is. I'm hoping to bring this kind of uplift home to help my wife and me manage our lives and achieve our goals.
I think this is a pretty obvious move. Our household (like most households our age, probably) already leans on tech pretty heavily. We keep track of many of our records in Google Docs and Sheets, organize information in Notion, and sometimes plan goals or household tasks in Trello (my wife is the CEO of our lives, and she has a passion for organizing information). We also enjoy talking to Amazon's Alexa, and we have several such devices in our apartments. Gmail is another big one: we probably email each other more than we text since we're together so often, and when we do need to send a written message it's usually to share a resource or note an idea.
We've built up a pretty decent corpus of messages, emails, and documents about our household, lives, and interests, and agents can use and analyze all of it to help us going forward.
I've recently been learning about Andrej Karpathy's LLM wiki pattern and think it could be a good way to boil down this household information into a maintainable, self-updating system. I'm not a huge Obsidian fan, but it has been helpful for seeing how the information gets organized and for spot-checking the raw files and ingestion flow. I used Claude Code on my MacBook Air to help create the initial vault and ingest the raw sources, and I back up an encrypted version of the vault to my personal GitHub.
There are a lot of great harnesses, many of them open source, that can be used as personal AI assistants for things like household management. I considered Claude Code, pi.dev, OpenClaw, OpenCode, and Hermes Agent, and ultimately chose Hermes after some experimentation, since it seems well suited to what we're trying to accomplish. It works with OpenRouter, which has a lot of cheap (or even free, shoutout to NVIDIA Nemotron Ultra) model options. It also has a really nice Discord bot integration, and my wife is a big Discord fan.
I originally developed this on my MacBook Air, which had to stay on for the Discord gateway server to run. After the expected amount of hammering out permissions and integrations with Notion, Google's API, Discord, and the LLM wiki's design and maintenance cycle, it started working quite well.
I've promoted it to a $12/month, 1 CPU, 2GB memory VPS on DigitalOcean so it can run 24/7, and I use my Claude Code session on my MacBook Air to SSH in for management. I could totally get a small Raspberry Pi to run this 24/7 in my apartment, but I don't want to make that kind of upfront investment right now given how quickly things keep changing in the AI world. For similar reasons, I'd rather not buy something like a Mac mini and run a smaller open-weight LLM to make token usage free. I can already use NVIDIA's free models, and the upfront cost of a new computer doesn't make sense to me when I only need to pay about $12/month for hosting and maybe $10/month for compute.
Lessons Learned
- I originally tried to create the LLM wiki vault with Hermes powered by OpenRouter models, but I kept getting greedy and creeping up the model and effort levels until I ended up spending a lot more than I would have with a one-month Claude Code Pro (or Max) subscription. This is what motivated me to build the "Tell Claude" flow into the Hermes agent, for when deeper research or automation flows need to be set up on the VPS through Discord.
- I should have spent more time learning about different ways to design the LLM wiki's structure and integration flow. There are a lot of decisions to make around how new information gets inserted into the vault, what deserves to be in our household vault, and what can be left out with just links to more detailed information. I think vault management and design will keep needing tweaks as I learn what works and what doesn't when integrating household-specific information into conversations.
Icons from Simple Icons (CC0) and selfh.st/icons (CC BY 4.0).