One private place for everything you know. Your AI can read it, and add to it.
Every conversation starts from nothing, and the things you actually know are scattered across a dozen apps — every one of them a company that can change its price, its terms, or its mind about your data. MacBrain fixes both. One private memory, on hardware you own, that any AI you trust can read from and write to. Nothing about it leaves your machine.
One SQLite file on your own machine. Your notes stay plain markdown in your own git repo — those files are canonical, the search index is derived. Delete the database and rebuild it from your notes. The reverse is not possible, on purpose.
Twenty-two tools over the Model Context Protocol, so any assistant you authorize can search your memory, read a page, and write one back — as a plain file in your own notes, never an entry you cannot open. Not a plugin for one vendor's chatbot — a memory all of them can use.
Run it on one machine and reach it from all of them, each with its own token and its own permissions. A read-only agent can search everything and change nothing.
Search is keywords and meaning at the same time: SQLite full-text for the exact words, a local embedding model for everything you phrased differently, and the two rankings fused into one. The model runs in‑process, on your hardware. Your memory is never uploaded.
And it fills itself as you work, because a memory you have to sit down and feed is one you will stop feeding. You finish something with an assistant and the assistant writes it up — it was there for all of it. A session you meant to write up and didn't is waiting the next time you sit down, not gone. Say it wasn't worth it and that is recorded too, with your reason, and it stops asking.
And the part that sounds like a limitation until you have lived with the alternative: the daemon itself runs no language model. Ever. Retrieval is deterministic. There is a model in the system and it runs every night — outside, read‑only. It reads your brain and hands you a digest: this link points at a page you never wrote, this page has nothing pointing at it, here is what changed. Suggestions, in a list. It cannot open the database and it has no path to your files. Nothing tidies your notes while you sleep. You do, when you decide to, and the record says it was you.
It has been running in production — mine — since 12 July 2026. As of this writing that brain holds 147 pages and 1,247 indexed passages. It also knows about six links pointing at pages that do not exist yet, and it will tell you which six. That is the whole idea: a memory that keeps track of what is missing, not just what is there.
Measured against the tool it replaced, on twenty fixed questions with human‑graded answers: nDCG@5 of 0.898 against 0.780, recall@5 of 0.873 against 0.760, ten queries won to five. The report and both sets of figures are kept with the code, including the parts that did not go the way the design expected. The test suite is 384 tests across 48 files, and it is green.
There is no signup button and no hosted version. You run it — on a small server, or just on your laptop. It wants Node 22 or 23. And if your notes are not files you can keep in git, the notes side of it has nothing to index.
Encrypted sync between devices is not built; today they share one server over your own network. And it is not open source yet. That happens when it is genuinely ready for someone else's machine, not when it would make a better headline.
Picture the ordinary version of this. You sit down, and the assistant already knows what you decided last month and why you rejected the other option. You stop being the integration layer between your own tools. And when someone asks where your knowledge actually lives, the answer is a file, on a machine, in your house.
When it is ready to share it will be open source, and it will be here. No waitlist tricks, no hype.