Lately I’ve been trying to make AI useful in my day-to-day, not just something I can demo once and forget about.
For me, the interesting part isn’t a clever one-off answer. It’s when the system starts to stick around — remembers context, helps with research, keeps notes organized, and quietly does a bit of work even when I’m not staring at the screen.
I was also looking for something I could continue from anywhere, including on mobile, without feeling too tied to one model or one company’s chat history. I wanted the notes, memory, and context to build up slowly in a way that stays accessible to me.
A few things that have been especially useful:
1. A second memory layer
I don’t want my “brain” to live inside one model. Notes, logs, and raw research go into Obsidian, and agents help turn that into cleaner wiki-style pages over time. A lot of this is inspired by the LLM-wiki / second-brain ideas people have been sharing recently, including Karpathy’s broader direction around LLM-driven knowledge bases. I like that this becomes a personal semantic layer that keeps growing, even if I swap tools later.
2. A running work diary
Short summaries of what I worked on, what changed, and what I need to revisit have been surprisingly useful. It helps preserve the “why” behind the work instead of losing it in chat threads.
3. Daily research without tab overload
I’ve been using Hermes more for recurring research: track a few topics, pull in updates, summarize them, and drop the useful bits somewhere I’ll actually see again.
4. Background workflows through Discord
This part feels very real to me. I have dedicated Discord channels where Hermes can send alerts, summaries, and progress updates. That continuity makes it feel more like a working system.
5. Home Assistant + home ops
On the home side, I like having Hermes close to Home Assistant and the homelab for practical things: cleaner alerts, useful nudges, automation tweaks, and remembering what I already debugged once.
6. Kanban / triage instead of pure chat
Some work just makes more sense as triage, routing, and follow-up instead of one giant prompt. That feels much closer to how real work actually happens.
7. Different models for different jobs
I’ve also stopped expecting one model to do everything perfectly. Some tasks need a stronger model; some just need something fast and cheap. Being able to route that intelligently matters.
What I like about this direction is that it compounds over time:
- the notes improve
- the memory improves
- the workflows improve
It’s still a bit messy and experimental, but it’s already more useful than my old “open 50 tabs and lose the thread” workflow.
I’ve also become a lot more opinionated about privacy and boundaries once memory, notes, alerts, and automation start connecting together.
I’ll share more concrete examples from the projects I’m working on in a follow-up post.
