Tasu: your knowledge and your files, in one chat.

Why a brain

You read and watch more than you can hold. A year later you keep the idea and lose the rest: who said it, when, and whether a later source changed it.

Three tools cover part of this today.

• A chat assistant answers from its training data: an average of the internet, frozen at a date.
• An AI-compiled wiki, the idea Andrej Karpathy calls an LLM wiki, compiles your files once into pages. When two sources disagree, both stay on the page.
• Notes keep what you wrote down, in the order you wrote it.

A brain compiles your sources into positions: one view per concept, rebuilt from every dated claim behind it. When a new source contradicts a position, the brain asks you before it changes anything.

Three choices follow from that, and the rest of this paper is how they hold:

1. Compile once, read many. A source is read at the door. Every question after that reads the compiled concepts.
2. Sign every claim. Each piece of evidence carries its author, its date and its source.
3. You rule on clashes. A contradiction waits for your call, and every view survives with its date.

The model

Seven things make up a workspace.

• Workspace. Holds folders, assets and chats. It sees nothing of the other workspaces.
• Category. A name such as Wealth or Business. Each folder and each asset sits in one, or in none. One click on a category asks all it holds.
• Folder. A subject such as Health, a person whose views are dated so drift shows, or a personal folder that keeps your own words. Each folder has a one-line scope.
• Concept. One idea with one position. It holds a summary line, its evidence, its data points, its open clashes and its links.
• Evidence. One claim: its date, its author, the claim in a sentence, and the source it came from.
• Source. A link, a title, an author, a date and the brains it fed. Its note keeps the extraction: topics, figures and exact quotes. The raw text stays out.
• Asset. A document, a table, a page or a web page you keep whole. A chat reads it and proposes changes. It feeds no concept.

Every concept also has a card: a 1 KB copy with its title, its summary line, the first 300 characters of its position, its counts and its links. Searching, planning and linking read cards. A full concept runs 1.5 KB to 19 KB, and it opens only when a step needs it whole.

Drop

A drop takes a link, a PDF, a Word file, a text file, a video or a pasted transcript. Files are read in your browser, and only their text travels.

1. Check

   The link is reduced to its host and path, or to its video id for any YouTube format, and looked up in an index. A match stops the drop and offers to file it into another brain instead.

   Index lookup, no model
2. Fetch

   A web page or a video transcript is fetched once. A page is fetched over https, with 5 redirects at most, in 25 seconds, up to 60,000 characters. What was fetched is discarded after reading.

   No model
3. Read

   The text is cut into 18,000-character passes, 3 read at once. Each pass extracts every topic with its figures, names, dates, reasoning and exact quotes.

   1 model call per pass
4. Plan

   The extraction is compared against the brain's cards: up to 1,000 listed, within 60,000 characters. Each idea matches a concept or proposes a new one. Findings sort into new, echo and clash.

   1 call per 25 topics, 3 at once
5. Your call

   You see where it goes and every clash, flips first. Keep both, which is the default, new wins, mine wins, or reject the source. A drop needs a named author.

   You decide
6. Settle

   Each touched position is rewritten from its whole evidence list, plus this source and your ruling. Evidence keeps the newest 12 entries and folds older ones into one dated line.

   1 call per 8 concepts, 3 at once
7. Link

   In the background, each concept the drop touched gets up to 4 links to related concepts, in any brain. A free shortlist comes first: rare shared words, one title naming another, a shared source.

   1 call per 30 concepts

The drop ends with a receipt: the source id, the brains it fed, and the counts of positions rewritten, new ideas, echoes, flips and clashes kept.

Ask

A question runs in two model calls: one picks the concepts, one writes the answer.

1. Route

   A model reads a numbered index of concept titles, up to 80,000 characters, which fits about 2,300 titles. It returns up to 30 picks and up to 12 English search terms. Past 2,300 titles, the ones sharing words with the question go in first.

   1 model call
2. Rank

   The picks come first, then up to 10 linked neighbours, then the concepts matching the question's words. A title hit weighs 3 and a body hit weighs 1.

   No model
3. Open

   At most 30 concepts open in full, within 60,000 characters, each capped at 8,000. Up to 120 more are named by title, so the answer knows what else the brain holds.

   No model
4. Answer

   At the level you pick. Normal gives the answer with its figures. Educational teaches it from zero. Learning gives the steps and lets you reach it yourself.

   1 model call

Every answer follows one shape:

• The first sentence answers the question. Figures and dates stay inside the answer.
• Names go on one last line: Sources: author, date · author, date. A person brain names that person throughout.
• Newer evidence wins on the same question, and better data overrides that. An open clash is stated when it changes what you would do.
• A brain fed by fewer than 10 sources says so in its first sentence.

You can also talk. A mic next to send writes your words with Whisper large-v3 turbo, in 27 languages, for about $0.04 an hour of talking. The browser's free voice input stays one tap away in Settings.

The brains are written in English. Answers come back in the language of the question, or in English when Settings says so. A question in another language still finds its concepts, because routing works on meaning and the search terms are translated.

A chat reads what you pick: folders and assets ticked in the side list or the chat bar, a whole category, or names typed after @, such as @Pricing, @cm-deal or @All Wealth.

Assets

An asset is a file you keep whole: a Word, PDF, text, Markdown, HTML, Excel or CSV file up to 30 MB, a web page read from its link, or a blank document, page or table. Files are read in your browser.

• Nothing changes without you. A chat proposes a change with the words before and after, or a new asset with what it will hold. Apply, Turn down, then Undo.
• Versions. A file dropped under a name an asset holds becomes its new version. The one before is kept for Undo, and keeps its category.
• A text, written out. A pitch, an email or a plan asked of an asset comes back whole in the answer, ready to copy.
• Up to 4 in one message, beside the folders ticked. A question on a 300-row table costs about $0.002 on the default model.

The personal folder

A third kind of folder, next to subjects and people. You talk to it, and it keeps what you said. Its own chat is its only reader.

• Dated notes. Each message becomes up to 3 notes, signed "You". A change of mind keeps the old view with its date.
• People. A card for each person you name: facts, history a year at a time, their people, and what is still open with them.
• An interview. 335 questions in 17 chapters, each under 30 seconds. It skips what your notes already answer.
• A daily log. 8 questions about the day, one at a time, each answer filed as dated notes.
• A twin. It answers as you, with your own rules, and a test of 5 situations compares its replies with yours.
• Your folders and assets. Tick folders beside it and it calls them. Name an asset and it fills it from your notes, as a proposal to apply.

Clashes

A brain that swallows every claim contradicts itself. So a brain argues back, at two moments.

At the drop

The plan sorts each contradiction into one of three kinds. All three reach the card, flips first, and keep both is the default.

After the drop

A clash found while rewriting is stored on its concept as two dated claims and the reason they clash. Before it is shown, a model checks it once, 40 clashes per call: it counts only when both claims cannot hold at once. The result is stored, so a clash is checked one time.

You settle them one card at a time. Swipe left and A holds, right and B holds, up and both hold, down to decide later, with 4 seconds to undo. When one side holds, the position is rewritten with the other kept as one dated sentence. When both hold, only the flag clears.

Why it stays fast

Two rules cap the cost of a question as a brain grows. Search reads cards, and whole concepts open only when picked. Every read has a ceiling.

The three reads of a question add up to 152,000 characters of your brain at most, whatever its size. A test suite holds this: 11 brains and 1,000 concepts, and the answer is found inside that budget.

Guarantees

• A position is rebuilt, not appended to. Each drop rewrites it from its whole evidence list, so a sixth source rewrites the view from all six.
• Every source is checked twice. By link at the door, then by idea in the plan.
• Every view survives with its date. A flip keeps the old view in the evidence.
• Every claim is signed. A drop with no named author stops before it files anything.
• Raw text stays outside. The note keeps the extraction and the exact quotes.
• Every drop ends coherent. The positions it touches are rewritten before the receipt prints.

Workspaces, keys and privacy

What comes out

• One-pagers. Folders, a category, up to 4 assets or a question become a summary, an 8-question quiz, a deep dive of 500 to 800 words, 3 to 5 use cases, or your own type. In English, French, Spanish, German, Italian, Portuguese, Dutch, Japanese or Chinese.
• A health score out of 10. Variety of authors weighs 3, depth weighs 3 (the share of concepts with 2 sources or more), freshness weighs 2 and open clashes weigh 2. Each brain is scored against your best one, with the one move that closes the gap.
• A connector for Claude. Through MCP, 7 tools read the brains and 6 more feed them on a signed address. The client's own model does the thinking. 120 calls per 10 minutes per address.
• A weekly digest. Mondays at 06:00 UTC: the sources read, new concepts, concepts fed again and open clashes.
• An export. One markdown file with every brain, concept, piece of evidence and source.
• Chats that stay. Saved on the server for 30 days after their last use, with 5 pins.

Stack and tests

How it writes

Everything a brain writes follows 9 rules. So does this paper.

1. W1No em-dashes.
2. W2Under 30 words per sentence.
3. W3Data replaces adjectives.
4. W4Weasel words removed.
5. W5Every line passes the "so what" test.
6. W6Simple expression.
7. W7Positive phrasing.
8. W8Quotes stay exact.
9. W9The brain writes in English. Quotes keep their language.