Skip to content

Home

Jump to bottom
Tunca Üçer edited this page Sep 28, 2026 · 21 revisions

brainless

Notes arrive from everywhere, collect in one place, link themselves to each other,
and come back to you when they are relevant. Then they argue with you.

Install  ·  Lite  ·  How it works  ·  Design  ·  Latest release

It runs on your own machine. A folder of plain text notes, a handful of small scripts and whichever AI model you point them at. No server, no account, nothing to sign up for. Nothing leaves the machine except the model calls you choose to make, and a local model keeps even those at home.

Sherlock had the method. Watson wrote it down, asked the obvious question, and kept the files. This is the Watson. You keep the method.

This wiki documents the public engine, ezapmar/brainless-public. The README is the one-page story; these pages are the detail.

The problem

Notes scattered across six places: voice notes in the car, screenshots on the phone, meeting transcripts in e-mail, articles you meant to return to. Written down somewhere and then effectively gone, because "I wrote that down" and "I can find it again" are not the same sentence. And underneath that, decisions made on a phone between two meetings and never looked at again, which is how you get to be wrong the same way twice without noticing.

A pile of notes does not fix it. Notes that nobody reads are a warehouse. Summaries that nobody argues with are a tidier warehouse. And an argument that nobody grades is entertainment.

The solution

One loop, and each part is there because the one before it is useless alone. Collect everything without judging it. Let the machine read it and make the connections. Argue with what you are about to decide. Write the decision down as a bet, and settle the bet later. Then keep the promises you made along the way.

The notes were a warehouse. What I wanted was a colleague.

Line What it means Pages
1. Gather information Everything comes in through one door, nothing judged on the way in. The machine reads it, summarises it and makes the links, so a note from last year turns up beside the one you are looking at today Capture Flow, Method, Scripts Reference
2. Think deeper and clearer Ask what you already know before anything hard. Then six personas, each built out of a book on thinking, argue the thesis in two rounds, and a flag goes up when they agree with you too often Personas and Dialectic, Thinking Clearer, Commands
3. Decide A decision is a bet: a prediction, a confidence and a review date. It never picks for you, and it nags until you write down what actually happened The Loop, Today Queue
4. Bonus: get shit done Promises from meetings and voice notes become tasks, without duplicates, synced with Google Tasks. At most three things each morning. Reminders you answer in a thread, and nothing is written until you say "apply" Today Queue, Buzz Interactions

The full walk through all of it, with the mechanics and the file names, is How It Works.

Start here

What it runs on

One laptop is enough, and that is what the installer sets up. A folder, whichever model you already use, and one scheduler entry that catches up after the lid was shut. I call it Lite, and it is where I would start.

The reference setup grew into three devices.

  • iPhone 16 Pro. Capture goes in through a Telegram bot (voice, photos, links, text). Everything that comes back (receipts, the morning three, the six voices, alerts) arrives as a Buzz channel on the same phone, and the reply goes in the thread.
  • MacBook Pro (macOS). Where notes are written and commands run, with Claude Code on a subscription, no API key. launchd runs the jobs while it is open.
  • A laptop running Omarchy (Arch Linux). The always-on worker, reachable only over Tailscale. systemd timers, the Telegram capture with local whisper.cpp, the self-hosted Buzz relay and the persona agents all run here, and it pulls, commits and pushes so the Mac can sleep.

The vault is one git repo between the two computers. The units, the schedule and the sandbox are in Reference Deployment.

Keep it yours

A loop that runs unattended against a folder holding your life needs one rule above all the others: nothing writes into your notes without you. The machine proposes and you apply. You own Work/, Personal/, Library/, Inbox/ and Thinking/; the machine owns the compiled layer, .wiki/. Folders you mark private are never read by the model. The rules are checks, not requests: hooks ask before a write into your folders or a deletion, and refuse secrets. Once a month the whole vault is sealed with a key the machine cannot open and put where sync does not reach. What you teach one agent, every agent knows: your standing preferences live in one file in the vault, not in one tool's memory. See Ownership and Privacy.

What it is not

  • Not a chat app, although it is getting conversational. The notes are still the record. The threads are how the record gets made.
  • Not a knowledge graph you maintain. The graph is built for you, not curated by you.
  • Not a decision maker. It never picks. Duke would call that resulting in advance.
  • Not finished. Built for one person, then made installable. Expect rough edges.

Where it is going now

For a year the system was a pipeline. Notes went in, a compiler read them, timers posted the results and the owner read them on the phone. It talked. Nobody talked back.

That changed on 21 September 2026. Telegram now only listens. Every answer, every question and every approval lives in a Buzz channel, in a thread, next to the message it belongs to. "Apply" under a preview writes the note. "Defer to Friday" moves it. Anything else gets an answer with the vault paths it stands on. Two channels go further: in #writing an agent drafts long-form pieces one thread per piece and runs the editing lint before showing anything; in #narratives another writes children's stories chapter by chapter. Both read the whole vault and write to exactly one folder.

So the work now is workflows and conversation. The pipeline stays. What is being built on top of it is the set of bounded exchanges a person and a system can have without either of them making a mess. A question with a revision-bound approval. A draft with a lint report. A pitch with its sources. Each one is small on purpose, and which of them survive a month of daily use is an open question. See Buzz Interactions and Writing Agent.

On 23 September 2026 search learned meaning. It had matched words, and the questions had stopped sharing words with the pages: Turkish about an English book, a synonym, a role described without its name. A local embedding index now sits beside BM25, and on the golden questions the right page moved from 62% to 79% at the top. The same index finds pages that mean the same thing but never link, and a nightly pass called dreaming asks about five of them each morning in #dreaming. A yes writes the link and it survives every recompile; a no is never asked again. The wiki proposes its own connections, and a person still decides. See Scripts Reference.

Pages

brainless · not a second brain, zero brain. Collect everything, think it through, decide, then do it. Engine repo · MIT. This wiki mirrors the repo's docs/; the repo is the source of truth.

brainless

brainless

Home

The loop

1. Gather information

2. Think deeper and clearer

3. Decide

4. Get shit done

Keep it yours

Run it

Addons

Reference

Repo · README

Clone this wiki locally