Your documents

The documents you give your AIs live in one collection of your own: books, articles, letters, notes, code, whole folders of them. It is yours, not any one AI's. Each AI is granted the documents it may draw on, and a document given to two AIs is read and remembered once.

Everything here happens on your computer. A document is read, cut into passages, indexed and summarized by models running on your device. No page of it is sent anywhere unless you ask an online model a question and the passages that answer it ride along with that question.

Adding documents

Open an AI's Knowledge tab in its editor, or the Remembers tab of its Memory page, and drop files or folders on the panel, or choose them. A folder is read through, subfolders included. Reading runs one document at a time; a book takes a moment, a decade of articles takes a while, and you can keep working. The activity card at the bottom right follows you to every page with the document being read, the pieces remembered so far, and a Stop, and it names the AI the documents are for. The Knowledge tab and the memory page of that AI show the same line, and their rows appear as each document lands. Reading keeps going while the app is open; your AIs draw on each document as soon as it is in.

What it reads: PDF, Word, OpenDocument, EPUB, text and Markdown, spreadsheets, HTML, and code. Photos and scanned PDFs without a text layer are reported as unreadable rather than silently skipped. The app keeps its own copy of the text, so a document goes on answering even when its file is gone.

Documents are read whole. There is no cut at a set size, so the second half of a novel is as available as the first. A Markdown page from a notes app is taken in by thought, not by length: each top-level bullet stays together with its sub-points, a title and tags are kept as words while ids and other bookkeeping are dropped, [[links]] read as plain words, and a long passage is never cut in the middle of a word. A very short document shows its own words on its card.

Adding documents without the small helper download says so where you add them, with the download right there.

Keep a folder in sync

Drop a folder on an AI's documents, or press Keep a folder in sync - your Logseq or Obsidian notes, or any folder of documents - and it stays current: new files are read, edited ones are read again (the document keeps its card and who can read it), deleted ones leave. Checked a few minutes after you open the app, every half hour, and when you press Check now. A folder that cannot be read (a drive that is away) changes nothing, and Logseq's own backup copies are left out. The Obsidian and Logseq tools offer the same sync with one switch.

Inside a synced folder the folder decides: a file deleted there takes its document with it, whichever way the document first arrived. Outside a synced folder nothing is ever removed for you.

Files kept in a cloud drive are left alone. When OneDrive, iCloud or Dropbox keeps a file online only, the app sees the name and never opens it (opening would download it), never takes it for deleted, and says how many it skipped. This matters even if you never chose a cloud folder: with OneDrive's folder backup or iCloud's Desktop & Documents, your own Documents folder is one.

Documents stay linked to their files

A document you give an AI keeps track of its file. When the file changes it is read again, and the row says when. When the file cannot be found, the row says so and the document goes on answering from what was read, with three things to do on its row:

• Check again looks where the file was; back with the same words, the row clears with nothing read; still gone, it says so.
• Relink is for a file that is somewhere else now. A relinked file with the same words is only moved, not read again.
• Remove takes the document out.

If a whole folder moved (three or more documents from one place), one Locate finds them all again, and Search for them looks through your home folder and the usual places on this computer - a file counts only when its words match, never by name alone; an external or network drive is searched only when you choose it. A document that went offline and turns up in a synced folder is recognized by its words and pointed at the file, never added a second time beside itself; if its words are already in the library under a newer record, the two are merged and the older record, with its card and history, keeps the file.

The list looks for moved and edited files when you open it and when the app comes back to the front, at most once every ten minutes, and each row can be read again by hand. Files dropped on the chat input are one-message attachments and are not tracked.

The library does one job at a time - adding documents, a folder check, a search for missing files, finding files after a restore - and each running job has its own Stop. Both document lists have a tick box on each row and a Select all, with one Remove for the chosen ones after a confirmation.

What each document gets

• A card - a short summary written on your device: what the document is, what it covers, who it is for. It shows under the document on the Knowledge tab and the memory page.
• A Mine tag - marks writing that is yours. It is guessed from the file's own author field (Word, OpenDocument, PDF and EPUB files carry one) against your Flowsta name, your username and the name you have told an AI, and it stays yours to flip.
• The author shown beside the name, when the file names one.

Cards are written by the helper model or, without it, by the local model you have loaded. They are never written online.

How your AIs use them

When you ask a question, the AI is handed two things beside it: a one-line card for every document it has, so a question about the collection ("what do my notes have in common") sees all of them by name; and the passages that match the question, chosen by meaning across every document it has. Only passages that fit the question are given: a passage has to score close to the best match, and when nothing in your documents fits, no passages are given at all - the AI still knows which documents it has. A short follow-up ("and the updated figure?") is matched together with the question before it. Answers say which document they drew on, and the Sources button under the reply lists which documents, how many passages, and the passages themselves as the AI saw them - in an ordinary chat, when the AI is using a tool, and when it is working in a project. These are the passages the AI was given, not proof it used them; Verify sources still ties a claim to a quote.

On a local model the block is sized to the room the model really has, so a small model on a short context still gets what fits. If you have pinned a model's context in Fine-tune and a turn needs more, the reply says so rather than trimming quietly.

What your documents say about you

Once cards exist, the Your Memory page adds a second paragraph to what your AIs are told about you: what you write about, from the documents tagged Mine, and what you keep, from the rest. It is written on your device whenever a card or a tag changes, and shown there on request.

Backups and a new machine

Your Vault backup carries the records: which documents, their cards, their tags, where each file was, which AI may draw on each, and the folders you keep in sync, by their place (Documents, Downloads, the desktop) and path. It never carries the text, so a large collection can never be the reason a backup fails. After a restore, the app looks for each file where it was last seen, and under this machine's home folder when the old one differs, and reads the ones it finds straight back into their records with nothing asked; a synced folder is in sync again where that place exists. Only files that moved are asked for, with the same Relink, Locate and Remove as any document whose file is missing. See Backups and recovery.