_agent/learned/2026-07-02.md
The AI that shows its work.
Most AI learns about you in a black box.
Kempt's writes it down, as plain Markdown, in your vault.
Search it. Edit it. Delete it. It's a file.
--- area: freelance learned: 2026-07-02 --- Hana Studio pays net-30. Draft the invoice on the 25th. You plan Mondays. You plan Sundays, execute Mondays. edited by you, saved.
Watch it learn in real time: two lessons land in
_agent/learned/, and the second one is yours to fix.
_agent/learned/2026-07-02.md
Correct its memory.
Agents like Hermes learn plenty about you.
What they learned lives where you cannot reach it.
Kempt keeps every lesson in _agent/learned/, one dated Markdown note per lesson.
When it gets you wrong, strike the line and write the truth.
The correction becomes the memory, and every plan after that obeys you, not the guess.
That is the whole meaning of observable: memory you can open, search, and overrule.
11:04 pm → 7:00 am
From braindump to done.
You dump the mess. The AI drafts the plan.
You check the boxes.
Every handoff is visible, in plain files you can open.
1 · you dump
invoice hana studio
poster board for Mina by fri
ship launch notes
call the venue before friday
2 · it drafts
Tue, Jul 2 · 4 tasks created
- Invoice Hana StudioHS-4.md
- Poster board for Mina by FridaySF-3.md
- Ship launch notesLN-7.md
- Call the venue before FridaySF-5.md
drafted by AI
3 · you finish
- done: Invoice Hana Studio
- done: Poster board for Mina by Friday
- Ship launch notes
- done: Call the venue before Friday
clients/hana-studio.md
Watch every edit land.
The AI panel shows its work as it happens. Tool cards you can open, and every file edit as a real diff: one line removed, one line added, on a file in your vault.
Locked your Tuesday around the Hana Studio invoice. Updated the client note so the net-30 timing holds.
Trust is a dial.
Five permission modes, from ask-first to full auto. Shift+Tab cycles them mid-session, and the riskiest ships locked behind a master switch.
Ask current
Asks before every edit.
Edit
Auto-approves file edits.
Plan
Plans only. No writes.
Auto
Runs on its own.
Bypass
Skips every check. Off until you unlock it in Settings.
Shift+Tab cycles the modes mid-session
input details
In Ask mode this card appears before anything touches your files.
schedule
It works while you are away.
Set a routine and link the notes it needs with [[ ]].
Every run, the AI re-reads those files, so a monthly invoice always has this month's context.
Leave your desk, come home to a draft.
Turn a chat into a skill.
One button reads the conversation and keeps what is worth repeating as a skill note, scoped to the right area. If there is nothing worth keeping, it writes nothing.
Plan my Tuesday
Locked your Tuesday around the Hana Studio invoice.
+ Pays net-30. Draft the invoice on the 25th.
Learn from this session
Learn to
- Global
- freelance suggested
- family
- side-project
_agent/skills/invoicing-hana-studio.md
Invoicing Hana Studio
Draft the net-30 invoice on the 25th.
AI-learnedPending review
New skills wait as pending until you approve them. The learning runs in its own session, so your chat stays untouched.
_agent/skills/
It leaves the receipts.
Everything it learns lands as a skill note in your vault, grouped by the life it belongs to.
Open it, edit it, approve it, or delete it.
- Weekly planning rhythm Plan on Sundays, execute on Mondays. AI-learned Active
- Invoicing Hana Studio Draft the net-30 invoice on the 25th. AI-learned 2 files Active_agent/skills/invoicing-hana-studio.md
# Invoicing Hana Studio Hana Studio pays net-30. Draft the invoice on the 25th. 1. Pull the month's hours from clients/hana-studio.md 2. Itemize from tuesday.md and the week's work notes 3. Draft, apply net-30 terms, confirm before sending
- templates/invoice.md1.2 KB
- templates/line-items.md0.4 KB
- Weekly client status update Summarize progress from tuesday.md every Friday. AI-learned Pending Approve
- School festival prep Track sign-ups and supplies from family/school-festival.md. AI-learned Active
No black box. Just files .
Provider neutral. Claude by default.
The AI is a guest in your vault, never the landlord.
Swap it, limit it, or switch it off.
claude by default
Claude Code ships wired in. Sign in once and it plans with the full context of your vault.
any AI, one memory Beta
Codex works today, more adapters coming. Skills and memory live as plain notes, so nothing it learned is trapped in one provider.
off switch in settings
Turn the AI off and Kempt stays a complete workspace. Local notes, plans, and boards, no cloud required.
dispatch · planned
Planned
Wake it from your phone.
Dispatch hands a job to the AI on your home desktop while you are out. It does the work at home, in the open, on the same observable surface you already audit.
Every run is visible line by line on the desktop. Nothing hidden.
Everything in the ai layer.
The complete catalog, with honest pills on what is still to come.
the full ai layer
Claude, wired in
The panel runs Claude Code on your machine with your own login, no API key injected.
Observable skills
It saves what it learns as skill notes in _agent/skills that you can read, edit, or delete.
Editable memory
What it remembers about you is Markdown you can correct instead of a black box.
Learn from a chat
One button turns a conversation into a reusable skill, scoped to the right area.
Safe write tools
It changes your work through narrow tools that enforce valid states and dates.
Live context
It sees the area, task, and note you are working on, and follows you as you move.
Scheduled runs
Set a routine and it starts the AI on its own, at a chosen time, with its own scope.
Permission modes
Cycle it from asking before every edit to full auto with one shortcut.
Off switch
Turn the AI off and the whole workspace stays complete and offline.
More providers Beta
Claude by default, Codex working today, and the panel injects no keys of its own.
Dispatch Planned
Wake your desktop AI from your phone and come home to green checks.
Rewind Planned
Roll back the Markdown files the AI changed, not just the conversation.