Memory is files.
Open them in any editor.
What VectorBrain remembers is Markdown in named folders on your disk. One search reaches all of it, by keyword and by meaning, and the keyword half works before you have a key.
VectorBrain’s real Knowledge panel in graph view, running on example notes. Every dot is a file or a row the app already has; the graph stores nothing.
Four homes, told apart by when they load.
The brief rides every turn of its project, so it has a meter and a cap of 8,000 characters. The wiki and the library are pulled only when a search finds them. There is no notes table behind the files, so what you delete in your file manager is gone from the app too.
Say “keep that”. It becomes a line in a file.
The assistant writes with memory.promote, which shares its writer with your Keep button. It lands as a journal row with an undo, and the brief it wrote to is a file you can reword by hand.
We never deploy on a Friday. Keep that in the brief.
A quiet conversation is read once. You decide.
By default, 25 minutes after a conversation goes quiet, the librarian reads it once and proposes notes of up to 800 characters, shown word for word. It can propose. Only you can press Keep, and a project’s librarian.md can narrow what it may propose, never widen it.
Searched two ways, labelled by which one found it.
Every question runs as a keyword search and a vector search, and the two rankings are fused. Each passage says whether meaning, keyword or both found it, because both agreeing is the strongest signal there is, and the only one you can check by eye.
hero direction
No key yet. Keyword search still answers.
Indexing without a key still reads, chunks and keyword-indexes the project, and says in a sentence why there are no vectors. A search then finds what shares a word with your question and misses what only shares its meaning.
hero direction
It knows your name because you wrote it down.
Your first name and a paragraph on how to talk to you ride every turn, capped at 80 and 600 characters. Your address and how you sign things go on about-me.md in the library, found by search when a turn needs them and never loaded otherwise.
An idea becomes a project and brings its things.
Promote an idea and the app gathers what shares its words: other ideas, files on the shelf, conversations and loose generations, up to 12 of each, with no model call. Untick what does not belong. One approval makes the folder, writes the brief and moves the rest in.
- truncate a brief
- keep a note you did not agree to
- load the wiki on every turn
- index a lockfile
- index your skills
A brief over its cap is not loaded at all, and the app says so, because half a brief reads exactly like a whole one. Skills stay out of the index because a prompt about writing would outrank your own notes about writing. How skills are found instead.
The knowledge layer, in numbers.
- Brief cap
- 8,000 characters, loaded on every turn of its project. Over the cap it is refused, not cut.
vb-cmd-memory/src/home.rs - Design sheet cap
- 8,000 characters, loaded whenever the design playbook is.
vb-cmd-memory/src/home.rs · vb-agent/src/run.rs - Where notes live
- <project>/.vectorbrain/ for the brief and wiki; ~/VectorBrain/library/ and ~/VectorBrain/ideas/ for the rest.
vb-cmd-memory/src/home.rs - Librarian
- Reads a conversation after 25 quiet minutes, at most 3 per 2-minute tick. Default model google/gemini-3.7-flash.
vb-agent/src/librarian.rs - Proposed note
- Up to 800 characters, kept only by a person.
vb-cmd-memory/src/lib.rs - Passages
- About 400 tokens each, with 60 tokens of overlap so a fact on a boundary is found from either side.
vb-index/src/chunk.rs - Ranking
- Reciprocal rank fusion, K = 60, over the top 30 of each search.
vb-index/src/retrieve.rs - Results
- 8 passages by default; at most 3 from any home other than the project you are in.
vb-index/src/commands.rs - Embeddings
- openai/text-embedding-3-small through your OpenRouter key, 1,536 floats. Any embedding model can replace it.
vb-core/src/embedding.rs - Largest file indexed
- 1,000,000 bytes. Lockfiles, .min.js, .map and .lock files are never indexed, and neither are exports or generated media: they are copies.
vb-index/src/ingest.rs - Watcher
- Re-indexes after 3 quiet seconds, at most once per 30 seconds per project, for folders talked about in the last 7 days and, always, the library, ideas and agents’ notes.
vb-index/src/watch.rs - You
- First name up to 80 characters and a style paragraph up to 600, both on every turn.
ui/src/You.tsx
The details,
for the careful.
Where does VectorBrain keep what it remembers?
In Markdown files on your disk. A project’s brief and wiki are in its .vectorbrain folder, your library is ~/VectorBrain/library, and ideas are in ~/VectorBrain/ideas. There is no separate notes database, so you can edit, move or delete any of it with your own editor or file manager.
Can I edit the memory myself?
Yes. Every note is a plain file. A folder talked about in the last seven days is watched, and the library, ideas and agents’ notes are always watched, so an edit is re-indexed within about 30 seconds. A brief edited past 8,000 characters is not loaded until it is back under the cap.
Does search work without an API key?
Yes, by keyword. Indexing with no key still reads and keyword-indexes the project and reports that nothing was embedded. Searching by meaning needs an embedding model, which runs through your own OpenRouter key. How the key works
Does indexing cost money?
Embedding does, on your own key, once per passage. After the first pass only changed files are read again, so a sync over an untouched project is a walk and a size-and-date check per file. Keyword indexing costs nothing.
Is it one memory across projects?
One search covers it. A search from a project reaches that project, your library, your ideas, the notes of the agent you are talking to and your businesses, and takes at most three passages from each home other than the project, so the project you are in stays first.
Who decides what gets remembered?
You. The assistant can keep something when you ask, through the same command as your Keep button. The librarian that reads quiet conversations can only propose; keeping and discarding its proposals are commands only a person can run. How the app separates you from a run
Can I turn the librarian off?
Yes, per project. A line reading sweep: off in the project’s librarian.md stops its conversations being read, and a homes line limits which homes may receive proposals. The file can only narrow what the librarian does.
What happens if I change the embedding model?
The app says first how many passages are embedded under the current model, because vectors from two models cannot be compared. Those passages stop answering by meaning until they are embedded again on your key. Keyword search keeps working throughout.
Can I export documents?
Your notes are already Markdown files. Documents made in the app export to PDF, DOCX and PPTX. Documents and export
Nothing it knows is hidden.
It is all in a folder.
If the assistant knows something about your project, there is a file that says so, and you can open it.