External Brain · v2.24.0
External Brain is the memory behind your AI assistants. It keeps the lessons from your finished work and your team’s decisions, brings the right ones back before the next task, and answers “why did we do that?” with sources.
It learns by itself in coding tools; in other work, from the lessons and decisions your assistant hands over.
Works withClaude Code · Claude Desktop · Cursor · Windsurf + any MCP client
Not a developer? See what Brain does for everyday work
Self-host with one command
$ git clone https://github.com/bejranonda/ExternalBrain.git && cd ExternalBrain && cp .env.example .env && ./scripts/dev-up.shOpen source (MIT). Works alongside the tool you already use. The hosted Brain is invite-only; self-hosting is free. Have an invite? Use your voucher
Works with the tools you already use
One installer per tool. Restart the tool afterwards so it loads the connection.
Once without Brain, once with it. Then one question about past work. Everyday examples, one after another.
Website form
YouAdd a contact form to my website.
AISure. Messages will go to the old info@ address.
AIDone. I added a blue Send button.
✗ You check it: wrong address — we switched to hello@ in August. And every form needs the privacy checkbox.
YouAdd a contact form to my website.
Brain remembers 3 things for this task
• Messages go to the new hello@ address — your decision, August
• Every form asks for privacy consent — this website
• Buttons use your brand green — learned on another project
AIDone: sends to hello@, with a privacy checkbox and a green button.
✓ You check it: right the first time.
YouWhy do our forms need a privacy checkbox?
BrainYou decided in August, after a customer asked how their data is used [1]. It is saved as a rule, so it applies to all your projects [2].
[1] Work session · company website · 14 Aug 2026
[2] Rule · “Every form asks for privacy consent”
Illustrative examples. In coding tools Brain also learns by itself; in other work it keeps what you or your assistant tell it, and your assistant answers from it.
Not only for code
Brain is not tied to coding tools. Connect it to a chat assistant on your desktop and it keeps the lessons and decisions you hand it — with the same honest labels.
At the end of a task your assistant hands over what it learned — “this client wants prices with VAT”. Brain keeps the lasting, specific ones and brings them back next time.
Say “remember: refunds within 30 days”. Brain keeps it as a decision, and your teammates’ assistants see it too.
Ask your assistant “what did we decide about the supplier?” — it looks it up in what Brain recorded instead of guessing.
Hand your assistant the notes. Brain keeps the decisions and open questions for the next meeting.
When a project ends, Brain learns how its documents were built, so the next one does not start from a blank page.
“What do I tell the board?” — a one-page draft from decisions and open items. Only when you ask; never sent on its own.
Switch on with V2_ORACLE_TASKS
Chat assistants that connect today:Claude Desktop
Good to know: Brain learns by itself only in coding tools — there it pulls the lessons out of each finished session. In other work it keeps what your assistant hands over: a decision you state (“remember this”), or the lessons worth keeping at the end of a task. Ask your assistant to do that: its built-in instructions talk about coding, and we have not yet measured how well Brain keeps lessons outside coding. Its built-in Ask page is tuned for coding questions.
The hosted Brain is invite-only while we grow it carefully. Write to us in a sentence or two about what you would use it for — we answer every request, but invites are limited.
Honest labels: what runs by default, what an operator switches on, and what we are building next.
When your tool starts a Session, Brain returns the Skills that match the task by meaning.
The Oracle answers in plain language and cites the Sessions and Skills behind every claim.
Your personal rules travel with you, so a fix learned in one repo helps in the next.
New Skills are suggested from patterns you repeat. You apply or reject each one.
Tag a project decision and every teammate’s next Session sees it. Rules can be promoted org-wide.
A background worker reads finished Sessions, extracts what generalises, and lets unused rules fade.
Paste a transcript; get decisions, owned action items and open questions to review.
Switch on with MEETING_UPLOAD_ENABLED
Items assigned to you appear when your next Session starts.
Switch on with V2_ACTION_ITEMS
The Oracle answers from your complete task list, not only from rules.
Switch on with V2_ORACLE_TASKS
Hand your coding tool a voucher; it creates the account, mints a token and connects itself.
Switch on with AGENTIC_ONBOARDING
The Oracle surfaces a relevant insight without waiting for a question.
See how rules and decisions connect across repositories.
Separate team vaults with access control and single sign-on.
Keep your Brain and your notes vault in step.
Four steps that run quietly while you code.
Your tool opens a Session and gets the rules your past work produced.
The outcome is recorded — what worked, and what you rejected.
Finished Sessions become typed Skills; the ones that pay off are reinforced.
Ask the Oracle and get an answer cited to real Sessions.
Built-in memory is useful. It just stays where it was made.
Sources, checked Oct 10, 2026:
Products change quickly; tell us if a line is out of date.
First reads from the project’s own validation record, with every sample size stated.
+0.148
better ranking of the right knowledge (NDCG@5) than plain similarity search
24 scored cases from real Sessions · 6 Jul 2026
4/6 → 6/6
tasks passing their tests with Brain knowledge vs without
n = 6, 0 regressions · 23 Jul 2026
3/5 → 5/5
on a second task suite
n = 5 · 28 Jul 2026 · uncurated re-run: 4/5 strict
Promising, not proven: samples of five and six tasks are first reads, not powered studies.
Checked Oct 20, 2026 against the public repository.
One command builds, migrates and starts the stack on your machine or server.
Mint a token, run the installer for your tool, restart it.
Decisions you tag reach your whole organisation. Team spaces are coming next.
Self-hosting is the main road. The hosted Brain is invite-only for now.
Needs Docker Engine 24+ and one LLM provider key.
git clone https://github.com/bejranonda/ExternalBrain.git external-brain
cd external-brain
cp .env.example .env # add one LLM provider key
./scripts/dev-up.sh # build · migrate · seed · startCreate a workspace on the hosted Brain, mint a token and connect your tool.
Use my voucherPaste one prompt into your tool and it installs the connection itself, then asks you to restart.
See the tutorialAlready have an account? Sign in
No. It doesn’t write code. It is the assistant behind the tool you already use: it remembers, carries lessons across projects and answers questions about your work.
Not to use it. Once someone sets up a Brain (on your own server) or gives you an invite, you connect it to your assistant and talk to it in plain words. Setting up your own server still needs someone technical. Outside coding, Brain keeps the lessons and decisions your assistant hands it.
Yes, with one difference. In coding tools it learns by itself: after each finished session it pulls out the lessons. In other work — writing, customer replies, planning — your assistant hands Brain the lessons at the end of a task, and Brain keeps the lasting, specific ones. Decisions you state (“remember: refunds within 30 days”) are kept either way. We have not yet measured how well it keeps lessons outside coding.
Built-in memory usually stays in one tool, on one machine, for one person. Brain is shared across tools, projects and teammates, and you can inspect and correct it. Use it alongside, not instead.
In your own Postgres, on your own server. Session text goes to the LLM provider you configure for extraction and answers; the privacy page lists exactly what leaves your instance.
Anything that speaks MCP. Ready-made installers cover the popular editors and command-line tools; others connect by hand.
Opt-in features are in the code but off until the operator switches them on. Coming next is on the public roadmap and not in the product yet.
Our app builder is paused for maintenance while we focus on Brain. Brain works on its own, with any tool.
More products from the same studio.
Is the water near me going up or down?
Live water levels of canals and rivers across Thailand, compared with the bank, with tested 12–72 hour trends. Built in Bangkok during the 2026 flood, in plain Thai, for phones.
1,646gauges nationwide
Is your home ready for floods, storms and quakes?
Ten-line “what to do” guides for floods, storms, earthquakes, fire, power cuts, heat and dust, every line with its source, plus a supplies check and a family plan card. The 60-second game is optional.
9“what to do” guides
Which shade of Thai politics are you?
A light-hearted quiz that maps your answers to one of 13 political “shades”, with emoji rain, 8-bit sounds and a running tally of everyone’s results.
13shades
Run the country, then see who you match
A policy simulation for Thailand’s 2026 general election. Choose real policies from 18 parties across six areas, watch the economy, welfare and democracy move, and see which party you match.
18parties
Build apps by describing them
A studio where you describe an idea in plain words and AI builds it with you, step by step. You see every change before it happens, and your Brain remembers your style for next time.
3languages
Self-host it today, or start with a voucher.
brain.autobahn.bot