Features

Vector Brain

From desktop 1.99.110, Vector keeps a brain for each repository and one for you: the rules you write, what you tell it about yourself, what it learns while it works, and the documents you give it. All of it is stored and searched on this computer. Beyond your rules, the start of MEMORY.md and the entries you pinned, which go with every request but a greeting, Vector sends your model only the few lines that bear on each request. Private documents never go to a free model labelled May train on your prompts, or to a session a guest is in.

Desktop 1.99.109 and earlier keep one MEMORY.md file, shown in Settings → General, and send it with every request.

What the brain holds

Rules: The repository's .vector/RULES.md and your own RULES.md in Vector's config folder, beside the AGENTS.md that Repository rules describes. Only you write rules; Vector never adds one on its own.
Your memory: MEMORY.md in the same folder, and what you ask Vector to remember about you. It applies in every repository.
What Vector learned: Short lessons, decisions and preferences the agent saved while it worked, each with the evidence it was saved from.
Documents: Specs, PDFs, runbooks and API descriptions you add, and the repository's own docs. See Documents and code links.
A repository's BRAIN.md: Memory a repository ships with it. Vector uses it once you trust the repository; until then it waits for your review.

Files you write stay yours: Vector only reads them. What it learns and what it reads from documents live in brain.db in Vector's data folder, beside its sessions. A key pasted into a memory or rules file is not sent to a provider, and Vector removes what looks like a secret from what it reads of a document before saving it. A copy kept with Keep copies of added files is the file as you added it, secrets included.

What a request carries

Every request but a greeting carries your rules, the start of MEMORY.md (about 800 tokens of it), the entries you pinned (about 300 tokens) and one line saying what else the brain holds. That part stays the same through a session, so the provider's prompt cache can reuse it. Then each request message recalls only what bears on it, within a budget set by the request's size, and recalls nothing when nothing does:

Chat: Nothing: a greeting recalls no memory and no documents.
Light: Up to 200 tokens, or 300 in a repository with documents.
Standard: Up to 300 tokens, or 500 in a repository with documents.
Thorough: Up to 500 tokens, or 900 in a repository with documents.

A greeting or a bare acknowledgement, such as "hi" or "ok", is sized Chat unless work is waiting on you: one that answers a question the agent asked, follows a run that did not finish, or approves a plan from Plan mode is sized as the work it carries on. A Chat request carries no memory, no memory instructions and no memory tool, so it costs no extra tokens. Reading a file brings in up to about 200 tokens of memory about that file, and the agent can search memory and open an entry with its memory tool when it decides to look. When a run nears its spend limit, a repository with documents recalls half as much.

Learning as it works

Vector saves a short lesson, decision or preference as it works: when you state one, when a fix makes a failing check pass, or when a lasting decision is made. It is in use from the next request. Each one shows as a Learned line in the session and under Recently learned in Settings → Memory, where one click undoes it; after Undo, Vector does not save the same words, or a close rewording, for 180 days. It saves at most three a request message.

What it learns in a session that read web pages or other outside content, ran MCP tools, or had a guest or a subagent in it waits for your review instead. Turn learning off with Let Vector learn as it works in Settings → Memory, or vector memory learning off.

Learning about you

From desktop 1.99.111: Vector learns from what you type: the formats and kinds of data you like, what you mean by words you use, and how you like things done. "When I say ship it, I mean merge and deploy" teaches a term, which comes into a later request only when that request uses the words. "I prefer tables", "always run the tests first" or "never force-push to main" teaches a preference. It reads only your own typed words, never text you paste, quote or put in code, and costs no model call. A short statement in a short message is in use at once; anything longer waits for a second sighting or one click. Vector tries not to keep sensitive personal details: a sentence that tells it about your health, money, religion, politics, relationships, sexuality or gender is dropped.

Each item shows under the reply as Learned from what you typed: "…", with Undo, and under Settings → Memory → About you, grouped by kind: Formats, Workflow, What you mean, Tools, Sources and Style. Each item shows the words it came from; edit, undo or forget it, or turn a kind off. Erase what Vector learned in this session forgets everything whose only evidence is that session. When your words teach nothing the rules can keep, such as a correction ("no, I meant…"), a small model can suggest what to learn: Let a model help Vector learn runs it with Vector Local only (the default), with any model, or never. Its tokens count in that session. Up to 300 entries about you stay in use; past that, tidying archives the ones used least.

Erase removes it from the brain and stops Vector using it. The sessions it came from keep their messages until you delete them. A model server you host may keep its own logs.

What proved useful

A card or memory entry that an answer cited, or that the agent opened, in a run whose checks went from failing to passing (or that the judge passed) gains weight. A record cited in a run you reverted loses it, and one given to the agent again and again without being used fades. The weight moves a candidate's value by at most 30% either way, and the counts behind it halve every 90 days, so it reorders what recall picks but never decides it alone. A session a guest is in, or that took in outside content, credits nothing.

Subagents and Vectorscope

In desktop 1.99.110, a subagent started by a Thorough request begins with up to 200 tokens of the cards and memory that bear on its task, for up to four subagents a request, and none for smaller requests. From desktop 1.99.111: every subagent also gets how you work, as Memory of you in every agent describes. With the next CLI release, a Vectorscope review you run in the terminal with vector review gets the requirements, constraints, contracts and decisions linked to the changed files, with any drift, in at most 300 tokens: private documents only when the review model is not a free model labelled May train on your prompts, never in a review posted to a pull request, and never content that waits for your review. Vectorscope reviews in the desktop app do not get them yet.

Memory of you in every agent

From desktop 1.99.111: every agent uses what Vector has learned about you: coding, subagents, the Vectorscope reviews you run in Vector and Deep research. Sessions with invited guests, and free models that may train on your prompts, get none of it. Reviews and research see how you work, not facts about you.

How you work is your preferences, decisions and lessons, the terms Vector learned and your own RULES.md. Facts about you are the facts you or Vector saved about you, MEMORY.md, private documents, and anything learned in a session that read a private document. For a role whose output leaves this computer, a decision or a lesson goes only when it names nothing in particular: no name, address, host, amount, date or long number.

Main agent: How you work and facts about you, as in desktop 1.99.110, for build and plan.
Subagents: Up to 100 tokens of how you work at every size, recalled with the task. A Thorough request's subagents also get up to 200 tokens of the cards and memory that bear on the task.
Vectorscope: Up to 300 tokens of how you review, chosen with the changed paths and file types, never with the pull request's own text. A pull request Vector cannot trust, such as one from a fork, gets your preferences, review notes and rules only. Vector tells reviewers that no memory line drops or downgrades a security finding or a blocking one.
Deep research: The research lead recalls how you work and the private documents its scope allows, beside what every request carries. Researchers get the formats, sources and style you prefer, and the citation judge the sources you prefer. A research run the build agent starts gets how you work and the private documents its scope allows, and nothing else private.
Titles, summaries and compaction: Nothing.

The lines each role gets come out of its own small budget, and a greeting still carries none. Reviews that run in GitHub Actions have no brain, because brain.db is on your computer.

Before Vector posts a pull request or a review, or makes a commit, it checks the text against what it knows about you. Text Vector or an agent wrote loses a sentence that repeats it, after you see what goes: post without it, post as is, or cancel. Text you typed or edited is never changed; Vector shows the overlap and asks. When the check cannot run, Vector asks whether to post anyway. Commands the agent runs itself, such as git commit -m or gh pr create, are not checked.

Settings → Memory

Memory lists this repository's entries or your own: search them, add one, edit, pin or forget it, and see what waits for your review. Recently learned lists what Vector saved on its own, newest first, with Undo. Tidy and erase merges duplicates, archives what goes unused and erases what Vector learned, leaving the files you wrote alone. Documents is described in Documents and code links.

From desktop 1.99.111: About you lists what Vector learned about you, as described above, and Research lists this repository's Deep research runs. Export and import saves what Vector keeps as Markdown to read or JSON to bring to another computer, for you, this repository or both. Imported items wait for your approval, and Approve all puts them in use in one click.

In the terminal

vector memory manages the same brain. The npm CLI 1.99.99 does not have it; it comes with the next CLI release, and Windows WSL servers get it once a later desktop release requires that CLI. Add --personal for your own memory.

vector memory list: Rules, saved preferences and what Vector learned. --learned lists what it put in use on its own, newest first, and --review what waits for you.
vector memory search <query>: Search memory the way the agent does.
vector memory show <id>: One entry with its evidence.
vector memory add <text>: Save something for Vector to remember.
vector memory edit <id> <text>: Change the text of an entry Vector learned.
vector memory pin <id>: Keep an entry in every request; unpin stops.
vector memory approve <id>: Let Vector use an entry that waits for your review; reject deletes it.
vector memory undo <id>: Take back what Vector learned; it will not learn it again.
vector memory forget <id>: Delete an entry for good.
vector memory tidy: Merge duplicates, drop what waited too long for review, and archive what goes unused.
vector memory learning on|off: See or set whether Vector adds to its memory as it works.
vector memory export: Print memory as Markdown: entries with their evidence, and what Vector read from documents.
vector memory erase: Erase what Vector learned in this repository; --personal for what it learned about you, --everywhere for all of it. Files you wrote are left alone.

Deleting your Vector account does not reach the brain, because it is on your computer. Erase what Vector learned and the documents it keeps with vector memory erase --everywhere --yes; the files you wrote stay until you delete them.