VEKTOR 2.0 is out of the workshop, and the list of what is new runs longer than we expected.
Skim the highlights below, or read the technical section if you want to see how it works.
What we believe in:
Cutting Edge Privacy Enhancing Technology
Sovereign local data principals
Protecting our customers' anonymity
Real-time LLM trivia game
Why we build it this way
Our motto has not changed since day one: thinking memory, not retrieving.
VEKTOR is a memory and assistant layer that runs on your own computer, works with frontier and open models, and keeps your data private where you store it.
This is also a story about how much one small team can now do. A developer anywhere in the world, with an idea, a laptop and an internet connection, can now build serious software using a mix of open and closed models. We think that is good for everyone, and VEKTOR is our proof of concept.
We are a small, self-funded team. We do not have rows of engineers or rented GPU data center inference farms, and we are not trying to be everything to everyone. We are putting our best efforts forward where a small team can lead: a four-layer memory graph, a useful harness around 600+ models, transparency about what the software does, and privacy by design.
The code is our own. VEKTOR is written in JavaScript, about 170,000 lines of it, to fit our own design rules. The few open-source parts we ship, such as PDF.js for reading PDFs, ONNX Runtime and the Silero model for voice detection, are named in the package with their licenses.
Testing every panel
Before this release we went through every panel the way a race team goes through a lap, looking for lost milliseconds. Panels that list your tools now share one copy, long lists load in pages, and the slow paths we found were tuned. Even a test Library of 1,400 books and 237,000 passages now opens its settings page in approx. two seconds.
The place where we see the most room left is local inference. Running models on your own machine through Ollama is the slowest step in the chain, and the next gains there will come from smaller models, speculative decoding and smarter routing between local and cloud models. That is where our attention goes next.
What VEKTOR does and does not try to be
People will ask whether VEKTOR does everything the big agent assistants do. It does not, and it is not meant to. It is a focused business tool: it remembers, it acts on your behalf with your permission, and it shows its work. We would rather do those things well, in the open, than promise everything.
Choice and privacy
We think the choice of AI tools belongs to the people who use them, in the same way you choose where your files live and who can read them. Free services are often paid for with your data. We took a different route: you subscribe to the software at a reasonable price, and your memory stays in a file on your own computer, complete with data sovereignty.
If you choose a cloud model, your prompts go to that provider, and we say so in the app. If you choose a local model, nothing leaves your machine. The decision is always visible and always yours to choose in the settings.
Why should you trust us? Because you can check our policies. The rules we hold ourselves to are written in our tenets, our privacy statement and the code itself, and the built-in action log shows what the software ran, what it asked about and what it was stopped from doing.
Technical highlights
Memory: a SQLite file on your computer, with four linked layers (semantic, causal, temporal and entity). Recall combines meaning, keywords and names, and now also scores single sentences, so a memory is found when the answer is one line inside it. In our tests retrieval takes single-digit milliseconds on a file of about 2,300 memories (retrieval only, not the model’s answer).
Changed facts: when two memories give different values for the same thing, they are listed with Replace and Undo. A fact restated in very different words is flagged as a possible change for you to review. Nothing is merged or hidden for you.
Voice, all local if you want it: speech to text with Whisper on your machine, read-aloud with Kokoro, and a small voice-detection model (Silero, about 14 MB) that runs in your browser to tell speech from a cough or a door.
Pictures: Vek tries your chosen provider first, then a picture-capable model on your own computer, then your other cloud keys. A key that has just failed is skipped for 15 minutes.
Permissions: an MCP tool asks before it runs unless its server marked it read-only. Scheduled routines pause after ten failures in a row, and one switch stops everything that would leave your computer.
Local-only server: the app runs on your machine, locks its window with a content security policy, and checks where each request comes from before it answers.
Quality gate: every build passes a 63-check release gate, and the packaged build is tested end to end against a copy of real data before we ship it.
What is new in 2.0
Vek, the assistant
Talk to Vek hands-free with Live voice, with turn detection, interruption, and a voice detector that ignores coughs and background noise.
Show Vek your camera or screen with Look: one picture, only when you ask, never saved. (Live Beta, with future improvements)
Four buttons under the ask box: Live, Camera, Screen and Chat.
Vek cards: checklists, step guides, option tiles and numbered lists, with icons, your own pictures, optional Wikipedia photos and small diagrams.
Vek opens Library books by title, at a chapter or a page, and opens any of 17 ready-made templates by name.
Vek reads the page you are reading and answers from it.
Email from Vek’s own address, with your approval on every send, plus an optional read-only mailbox from senders you trust.
Talk to Vek from your phone through Telegram.
Ask for a background task in one sentence, and Vek sets the name and schedule.
Eight mascots with their own voices and motion: the robot, a flat Orb with a colour picker, a glass orb with oscilloscope lines, Cloud, VEK 9000, Ghost, Game pad and the Magic 8 Ball.
Meetings and notes
Meeting notes in Jot: record your microphone and, if you wish, the other side of a call, or import an audio file.
Time-stamped transcripts with names, and notes from six templates: meeting summary, stand-up, client call, interview, brainstorm and lecture.
Calendar join: read an .ics file to fill in the meeting title and invitees.
Remember button to save the notes to memory, only when you press it.
Desk can save a decision when you state one, and Workflows keeps a Decisions list with an as-of date.
Library and reading
PDFs open in the same reader as books, with selectable text, highlights, zoom and find.
Library starter pack of 14 reference files, an index button for new files and another for a full rebuild, text editing with kept versions, and a Remove from reading action.
Convert tool with 81 conversions, a built-in guide of 63 chapters, and a Desk bar with Library search and Convert one click away.
Workflows and business tools
Projects: a goal, running notes and the result of each run, with the last plan reused on the next run.
Goals with steps that follow your task list, and optional steps suggested by Vek, labelled as Vek’s.
Ideas tab of things Vek can really do today, an Inbox tab that turns a pasted list into tasks, and Recipes for known code fixes applied the same way every time.
Triggers: run a Workflow on a schedule, when files in a folder change, when a commit lands, or when a CI run fails. Steer a run while it works.
Task due dates, reminders and a to-do list you manage by typing in Desk.
Companions page for Via, Vek-Sync, Prov and Vex, each with a preview before any action.
Flowire opens a file and lights up the files a change touches.
Memory
Causal links updated between memories, drawn as arrows in the graph.
Insights (beta, off by default): short summaries of related memories, kept apart and labelled.
Optional key facts: three short facts per Desk answer, off by default.
Real dates in place of “yesterday” and “last Friday” when memory is handed to the model.
Daily limits for web search, and Firecrawl and Keenable search keys.
Security and transparency
Action log: one list of what VEKTOR ran, asked about or was stopped from doing, with the rule that applied.
Pause all outside messages, a single switch for Telegram, email and delivery.
A safety message instead of a coaching run when someone writes about self-harm.
For play and for focus
Trivia on the Game pad: 8 topics, 3 levels, 20 questions written fresh by your own model, a timer, streak scoring, sounds and confetti.
Magic 8 Ball with a real 20-sided die and a fair draw of the 20 classic answers.
Ghost, Cloud, Game pad, Orb and VEK 9000 each have their own funny quips.
Models and skills
Seven more newly released models in the OpenRouter list, which still refreshes itself.
Two new skills: a plan-build-verify loop and a minimax-matrix for choosing between options on accuracy and speed.
What happens next
We will keep adding functions as we get feedback. Chrome automation, enterprise controls, more games and a game-pack format, and faster local inference are on the list for 2.1.
Try it, tell us what works and what does not, and hold us to the standard above.
Run your own race, and keep your sovereign data private and local.
Upgrading
Drop-in from any earlier version. Your memory database is untouched, and v2.0.0 includes everything from v1.9.10
The full changelog is at vektormemory.com/docs/changelog. The forum is the fastest way to reach us with questions or feedback.
VEKTOR Memory builds local-first persistent memory infrastructure for AI agents. Documentation and downloads are at vektormemory.com.
Top comments (0)