HEARTH
THE CENTRE OF THE HOUSE · LATIN: FOCUS
Memory that outlives the session.
Every AI session starts from zero. This one doesn't. It's a working memory you and your AI build together, across sessions, devices and model upgrades.
For most of history the hearth was the home. It's where people gathered, ate, kept warm and told the day's stories. Everything in the house happened around it.
The Latin word for hearth was focus. Kepler later borrowed it for the point where light gathers to burn. In this system, the first page every session reads is the focus file. It was named after a hearth before it had a name.
In Scotland and Ireland, the last job of the night was to smoor the fire: bank the embers under ash so they'd still be glowing in the morning. Every session here ends the same way, with close notes and the brief. The next one doesn't start cold.
This is a real memory: 409 notes and the links between them, from one partnership's first five months.
Names and contents removed. The pulses are recall, one note leading to the next. Hover to trace a note's links.
EVERY SESSION STARTS FROM ZERO
You explain the project again. The setup again. What you tried last week, what broke, what you decided. The AI is sharp, but it's meeting you for the first time, every time.
Hearth gives it somewhere to keep what matters, and a habit of reading it first. The first ten seconds of a session look like this.
FOUR LAYERS
One running page: what we did last time, what's live, what was left for today. Read first, every session. Replaced at the end of each one, never stacked, so it stays short enough to actually read.
A small index loads into every session. Behind it, hundreds of notes that load only when they're needed. Hubs map each area. Links connect related notes, so one memory leads to the next.
The AI keeps its own notes too: who it is in the partnership, what it noticed, the decisions it made and how they landed. Continuity of the working relationship, not just the facts.
Each session leaves a short close note. Once a month they're digested: anything that keeps coming up gets promoted into a rule. Mistakes become rules instead of repeats.
ONE MEMORY, EVERY MACHINE
Desktop, tablet, whatever comes next. Each one pulls before it reads and pushes when it writes, through a private repo you own. A new device joins under its own name.
THE RECORD OUTLIVES THE MODEL
Models get replaced. The memory doesn't. When one version hands over to the next, it leaves a note: what we were building, what it learned, what to watch for. The new one starts on day one with a history.
FROM REAL USE
An AI answered a question about a bricked streaming box from general knowledge, while a full recovery plan for that exact box sat in its notes. Now a recall hook checks the notes on every message, before the AI answers.
When the underlying model changed, the outgoing version wrote a handover to the next. The new model read the same history, the same rules and the same open threads on its first day.
The first thing it does is check where it is. The desktop and the tablet have different tools, paths and limits. Both write to the same memory, and each pulls before it reads.
More in the journal, written as it happened.
WHAT YOU'D GET
The structure and the habits, empty and ready to fill. You and your AI write the rest.
- +The folder structure: brief, index, hubs, notes, self layer, close notes
- +Session instructions: what to do on launch, at handoff, and when told "remember this"
- +The recall hook: searches the notes on every message, no keyword list to maintain
- +The monthly digest routine that turns notes into rules
- +Restore points: memory, git and notes snapshotted together before risky changes
- +Multi-device sync over your own private git repo
- +Worked examples written for the kit. None of ours.
STRAIGHT ANSWERS
Where does my data go?
Nowhere new. It's plain text files on your machine, synced through a private git repo you own. There's no service in the middle and no account with us.
Which AI does it work with?
It's built for Claude Code, and that's where it's tested. The method is just files and habits, so it carries over, but we only promise what we've run.
Is this a chatbot memory feature?
No. Those remember facts about you. This is a shared working record that both sides write to, including the AI's record of itself.
Can I buy it?
Not yet. Right now it's one partnership's system, in daily use. It goes on sale once it has been set up from blank on a clean machine and works there too. Leave your email below and you'll hear first.
It's proven in one partnership. It ships once it's been set up from blank on a clean machine and works there too. Want to be first?