My external brain
in production
Every AI chat starts from zero.
Mine doesn't.
🤖 Neo — the reasoner
OpenClaw · AWS EC2 · eu-north-1- Gateway + Telegram bridge
- Decides: answer or delegate
- Stateless per message, MCP tools
⚙️ Žaibas — the worker
Claude Code · Hostinger VPS- Polls GitHub issues for work
- Spawns Claude Code, opens a PR
- systemd + watchdog auto-restart
Twice a day,
I dump my head.
No format, no template, no discipline beyond showing up. The memory layer is only the storage — this is the source.
What I'm walking into. Yesterday's loose ends, what I'm dreading, what has to happen today.
What actually happened. What I finished, what I dodged, what changed my mind and why.
The system is exactly as good as what I feed it.
It remembers,
as flat files.
Markdown files + Voyage embeddings for semantic recall. No goldfish — context carries across every session.
- MEMORY.md the index — one fact per line
- daily/ notes what happened, when
- decisions.md why I chose what I chose
- projects.md what's in flight, and where it stalled
- contacts/ who's who · last contact · open threads
- "It reads yesterday's notes
before answering today."
Each skill is
a SKILL.md.
A procedure it follows — not compiled code. Adding a capability means writing a document.
17 cron jobs =
a nervous system.
It acts without a prompt. Proactive, not reactive — the value came from things nobody asked for.
"Scheduling beats reactive."
- 07:00 morning summary
- 21:00 nightly digest
- crypto + cost alerts
- therapy prep · weekly review · …
Every Tuesday,
it preps me for therapy.
05:30. It reads back my own week and hands me the things worth bringing — before I'm awake to ask.
- reads the reflection log + the last 7 days of daily notes
- pulls out patterns & tensions the emotional through-line of the week
- 3–5 bullets, no fluff on Telegram, 05:30 Tuesday
I wrote the schedule once. It hasn't missed a week since.
Tokens died overnight.
It fixed itself.
Zero human involvement — until the one OAuth click only I can do.
It reads the
internet overnight.
Scans AI + tech content while I sleep, rates each item by relevance to me, delivers a ranked briefing at 9am.
Not just data.
A decision.
Low recovery + a hard workout on the calendar → it connects the two, before I do.
- flags the conflict 🟡
- suggests cutting a set
- offers to message the coach
Three lessons.
- Context design > model capability the right memory beats a bigger model
- Scheduling beats reactive the wins came unprompted
- It's infrastructure now when it breaks, I feel it
The honest slide.
- Failures are silent a cron job reported "ok" for 9 days without ever running
- Tokens expire · prompts drift
- Model upgrades break things silently
- Cost is real — €405/mo 5× what I thought it was
€405 a month.
€4,860 a year. I told people $60–90 for months — I'd simply never added it up.
I pay €200/mo for a plan my agent can't use.
−€100/mo · cut in one afternoon, by auditing which model each cron job was quietly inheriting.
That's the system.
Ask me anything.
→ live demo (wifi permitting)
This slide
wasn't my idea.
One asked. One built it. I just forwarded the ticket.