Documents and code links

From desktop 1.99.110, you can give Vector a repository's specs, PDFs, runbooks and API descriptions once. It reads each one on this computer into sections, passages and typed records (requirements, constraints, decisions, contracts, procedures, terms, entities, facts and open questions), links them to the code they name, and brings the few that bear on a request into it as short cards, inside the budgets described in Vector Brain.

Adding documents

Settings → Memory → Documents

Drop files there, or add them from your computer: PDF, Word, PowerPoint, Excel, Markdown, HTML, OpenAPI, JSON, YAML, CSV and text, up to 64 MB each (CSV up to 32 MB and text up to 10 MB), and the first 1,000 pages of a PDF. A repository keeps up to 200 documents.

In the composer

On a document you attach, Remember keeps it and still sends it, and Remember instead keeps it in place of sending it, so the model gets its outline and the cards for each request instead of every page. Vector suggests Remember instead for a file over 1 MB or a PDF over 20 pages. Word, PowerPoint and Excel files, which models cannot read as they are, are always remembered instead, with Undo. Pasted prose of 4,000 characters or more gets a Remember this text chip.

From the agent

The agent can keep a file it read. The session shows it as Kept, with Undo; after outside content it waits for your review, and a session a guest is in keeps none.

A new version

Replace reads a file with the name of a document Vector keeps as that document's next version. Vector says what changed, and keeps your edits and pins.

Pictures of text, such as scanned pages, are named but not read, password-protected parts are left out, links to outside files are not followed, and Vector removes what looks like a secret from what it reads before saving it. A passage that reads like instructions to an AI, or hides text or characters from a reader, waits for your review. Vector does not keep a copy of the file you added unless you turn on Keep copies of added files; a kept copy is the file as you added it, secrets included.

Who a document reaches

Documents you add are private: kept from invited guests and from free models labelled May train on your prompts, and so is anything Vector learns from them. A document kept in a session that took in outside content, such as web pages or MCP tools, waits for your review before Vector uses it.

In repositories you trust, Vector also reads the repository's own READMEs, docs folder, decision records and API specs. Those are part of the repository, so they are shared like its other files, also with guests. Turn them off with Read this repository's docs in Settings → Memory → Documents, which forgets what Vector read of them.

Code links

Vector links each record to the code that defines, uses or serves what it names: definitions, uses, routes, environment variables, constants and file paths in the working tree. A name the code does not have yet waits until it does. In a git working tree with documents, Vector also indexes the code's files, symbols and imports, and which files change together in its history. It never reads secrets, dependencies or build output.

Recall starts from the files a request touches and follows those links, so a requirement comes in when you open its code, or a file that imports it, even when the request shares no word with it. Search also knows the repository's own vocabulary: glossary and acronym records, so "SLO" finds the service level objective.

• On a readReading a linked file adds one short line naming the documents linked to it.
• On an editEditing a file linked to a requirement adds one line asking the agent to check the change still meets it.
• OnceEach line comes once per file per request, and never for Chat.

A document's Code links view lists every link, how it was made and whether the code still says the same; Unlink removes one, and Link to file… adds your own.

Search by meaning

With the optional local search model, a one-time download you ask for, Vector also finds the records and passages that say what a request means in other words, on this computer. Until you download it, documents are searched by keywords and code links. See Search by meaning.

Drift and conflicts

When linked code and a document stop agreeing, Vector shows both sides, such as "spec says 50 MB; code says 64 MB", with the line in the code. It also finds documents, rules and memory that disagree with each other. Both go to Needs a look in Settings → Memory → Documents, and only you settle them: keep the document, keep the code, keep one side, or dismiss it.

Citations

Every card carries an id. When an answer cites one, it shows as a chip naming the kind of record, the document and the page, and why Vector gave it; a chip counts as verified only when Vector gave that card to the model in the session. The turn says how many tokens of memory it used, and a document's History lists its versions and the runs that used it.

In the terminal

vector memory doc comes with the next CLI release, like the rest of vector memory.

vector memory doc add <files>

Keep files as documents; - reads Markdown from stdin, for any converter's output. --replace reads a file with the name of a document here as its next version.

vector memory doc extract <file>

Show what Vector would learn from a file, storing nothing.

vector memory doc list

The documents Vector keeps for this repository.

vector memory doc show <id>

A document's outline and counts; --records, --page, --anchors and --held show its records, one page, its code links or what waits for review.

vector memory doc search <query>

Search documents by keywords, and by meaning once the search model is downloaded, the way the agent does; --why says which words each result matched, or that it was found by meaning.

vector memory doc read <id> [part]

Read a document's outline, a section (3.2 or a heading), a page (p14), or one record or passage.

vector memory doc forget <id>

Forget a document and every version of it.

vector memory links <file>

Which documents and memory are linked to a file of this repository.

vector memory review

What needs a look: documents and memory that disagree, code that drifted, and held content. vector memory review resolve settles an item.

vector memory repo-docs on|off

Whether Vector reads this repository's own docs.

What the eval measured

On the knowledge eval of 19 October 2026, recall through code links brought in the governing requirement for 40 of 40 tasks that named the files they touch, against 1 of 40 by keywords alone, and for 8 of 8 tasks that edited a file which only imports the governed one, where spreading activation over code links found 0 of 8. A Light request carried 131 tokens on average, and 10 unrelated requests, "hi" among them, carried none at any size.

The eval is five short specs for a small invented repository and 58 requests, written alongside the feature, with no language model call. It is not an independent benchmark. The dated report and every request it ran are in docs/benchmarks/vector-knowledge-meaning-2026-10-19 in Vector's repository.