DEV Community

Vektor Memory
Vektor Memory

Posted on

VEKTOR 2.0.6: Wan2.2 Image & Video generation—Suite: Deck, Folio & Docs tools

Today we’re excited to release VEKTOR 2.0.6. It lets you describe a picture or a short video and have your own computer make it. We also added an LLM assistant bar at the bottom of every Studio tab, so you can type changes and improve generation quality.

Suite is an addition to the most used business tools for spreadsheets, slides, and documents.

A complete job toolkit rebuild for resumes and cover letters that gives you options to create or extract info from a website or files to make creation faster.

Our Ethos in this build
Our ideas that shaped this release, and each one shows up in a feature you can use today.

You control your workflows. The assistant proposes; you decide. Edits arrive as a list you tick, memories wait in an inbox, deleting always asks, and every result can be undone.

Your data is yours. Memory, files and projects live on your computer, in files you can open and keep. Syncing between your own machines needs no cloud.

Your information stays private. Pictures, clips, spreadsheets and documents are made and converted on your computer. Secrets are refused before they reach memory, and Google sign-in asks for a single permission.

We do not have any watermarking in generations and will never have any ads in our software; all content is yours to use anywhere.

Security is built in, tools you connect are shown before they run, and anything that changes something asks first. Downloads are pinned to fixed versions and checked against a checksum.

Text from a web page or a file is treated as data, never as instructions, and the model can only return a closed list of edits that code checks twice. That limits what a prompt injection or a tool that changes behavior after you approve it (a rug pull) can do.

Cloud and local are always your choice. Use your own computer for private or free work, or switch to a cloud model when you need more power, the system notifies you of which model is being used.

We never record, save or use any of your data in our training or telemetry; we don’t have it to use.

Over 600 models and providers work with your own keys, plus Ollama local and cloud. Your plan also includes 20 million tokens of GLM-5.3-Flash through VEKTOR Included, with no key to paste, no account to make and nothing you wrote stored. A usage bar under the context window shows what is left.

Make pictures and short videos on your own computer
Studio > Image > Generate. Describe what you want, press Make it. Pick a style from one line (Photo, Ink wash, Watercolor, Claymation, Felt craft, Plush, Paper art, Pixar, Pixel art, Illustration), start from a Quick start (Portrait, Poster, Anime, Product, Landscape), and choose Landscape, Portrait or Square.

Enhance prompt turns a short idea into a detailed one with your own chat model. Afterwards you can make the picture better, upscale it, turn it into a short clip, or try another. A Recent row keeps your last ten.

The same panel makes pictures through a free provider (Cloudflare or Hugging Face), so the button works on day one.

On our testing rig, the RTX 3060/12Gb Vram, an image takes 12 seconds to generate with Wan2.2 included, an optional upscaler is also included. A five-second clip takes about five minutes, these generations will be much faster for more advanced spec GPUs.

**Animate your own pictures
**In Animate, pick AI video (on this computer) to move a picture of yours, or start from your words alone. You choose the length in seconds, and an estimate tells you how long it will take. Animate an image works on any picture, and the Shelf can send an image straight to Generate.

Use the royalty-free online sources to make images, animations, or videos from these options and remix as much as you like.

One assistant bar in every Studio tab
Type “shorten it to 20 seconds”, “dither this photo at size 2” or “add a total row”. Studio answers with a short list of changes. You tick the ones you want, and the sliders, pickers and cells move. Undo is one click. In Motion, the same bar makes a whole video from one line.

Simpler panels
Motion, Voice, Art, Dither and Generate now start with one box and one button. The adjustments open after you have a result, and the technical options stay folded under Advanced.

Edits and memories you approve
The assistant no longer saves things about you straight away. Memories wait in an inbox, the status bar shows “2 memories to review”, and you Keep, edit or Drop each one.

More tools, fewer places to look
Suite: Deck, Folio and Document move to their own button, so Studio is for making media and Suite is for business tools.

Sheet, Document and Plan: a spreadsheet, a document editor and a floor-plan tab, all driven by the bar.

177 more ready-made skills: 20 ship with VEKTOR and 157 are one click away in the Library.

Parts tab: import your own web pages and Studio splits them into building blocks you can search and reuse.

Dubbing and pronunciations: translate a script or subtitles into eight languages, and teach every voice in the app how to say your words.
Status-bar chip: Ready, Busy or Slow, with the reason, such as the model running on the CPU.

What problem does this solve for me as a user?
“I do not want to pay per picture or upload my work or have watermarks.” Generate runs on your graphics card. Nothing leaves your computer, there is no per-image fee, and a free provider covers you until you install the add-on.

“My graphics card is also running my chat model.” When a picture or clip needs the card, the chat model steps aside and returns by itself with its usual settings, so the next message is not slowed down. If memory still runs short, Free the graphics card unloads it for you.

“I do not know if my card is big enough.” The installer reads your card first. A card that is too small is told so before anything downloads, and only cards with about 10 GB or more are offered the finer-detail pack.

“The AI rewrote my whole document when I asked for one change.” The bar proposes small edits you can tick or skip, and everything can be undone.

“Simple requests take too long.” Totals, colours, formats, slide moves, a slower animation: these now run on your computer with no model and no wait.

“I need a spreadsheet, a document or a deck, and I do not want another subscription.” Sheet, Document and Design open your own Excel, Word and PowerPoint files and save Word, PDF and editable PowerPoint, all on your computer.

“I never know what the assistant saved about me.” You see every memory before it is stored, and you can edit it.

“Why is my local model suddenly slow?” The status chip names the cause. The common one, a model that has spilled from the graphics card to the CPU, was invisible before.

“I have to hunt for the right skill.” Desk picks a matching skill, and offers one from the Library when you have none that fits.

“I am afraid it will delete something.” Deleting a file or running a destructive command always asks first, even in Full access.

“My first message after a break is slow.” Desk’s small routing questions skip a local model that is not loaded yet. The rules decide at once, about 3 seconds saved, and the model warms in the background for the next message.

Technical specs: Generate: engine, packs and hardware tiers
The add-on runs Wan 2.2 through stable-diffusion.cpp (CUDA build) with GGUF weights. The installer downloads pinned files at fixed versions, checks size and SHA-256 on each, and resumes if interrupted. Downloads use a separate curl process, which reached about 20 MB/s on our line, against about 2 MB/s in-process.

The server reads your card once (nvidia-smi) and picks a tier: small (under 8 GB), mid (8 to 14 GB), large (14 to 20 GB), and extra large. The tier caps clip length, size and whether weights are offloaded. The pack guard runs before the download and again on the server, so a copy of the page cannot install a pack your card cannot run. A graphics-card lease unloads the chat model while a job runs and warms it back with the app's own settings.

Measured on one RTX 3060, one job at a time, chat model unloaded.

Known limits: the Large pack requires a 16 GB VRAM card for higher quality generations. a new model is due for release soon, we will update once released in Hugging Face.

The Studio bar: closed edits, checked twice
The model is never allowed to change anything itself. It returns a closed list of edits (set this control to this value, replace this text with that text, add this row). The server checks each edit against the tab’s own rules, the page checks it again on Apply, and anything the tab does not recognise is dropped. A bad answer can waste a turn. It cannot write to your project.

We tested the alternative first. Asked to add a meeting room, the model returned the whole room list rewritten. In four tries it was correct none of the times: rooms dropped, seat counts changed. Showing it more of the plan (160, 800 and 2,000 characters) changed nothing, so the problem was the task, not what it could see. One find-and-replace fixed it.

One machine, one local model, one request at a time: a direction we decided. We rejected the shorter controls prompt because it was faster and worse. If the model answers in words before the JSON, the server retries once with “JSON only”.

Fast routing for small questions
Desk asks a model small closed questions (which mode, yes or no). Three measures keep them from slowing a chat:

Local only if loaded. A cold 10 GB model takes about 30 s to load, which trips the first-byte stall guard. The server checks /api/ps (a few milliseconds) and, when the model is not loaded, lets the rules and the embedding shortlist decide, then warms the model for next time.

Cache-friendly prompt. The router’s instructions are now identical on every call and the allowed modes and the message come last, so Ollama reuses its cache after the first call. On a loaded model, routing answers took 282 to 1,537 ms across seven different messages, against a 3 s limit.

A fail-fast waterfall for small calls (opt in per call): providers are tried with a deadline, a rate-limited or refused provider is benched for the time it asks, a slow leader is hedged by the next provider, and an answer that fails its check is passed over. It follows the cooldown and hedged-request patterns used in “The Tail at Scale”.

Sheet, Document and Plan
Sheet: formulas (=F10-D10, =SUM(B2:B5), AVERAGE, MIN, MAX, COUNT, ROUND, ABS, IF), colours, percent and money formats, sort, and charts (columns, bars, lines, areas, stacked, pie, donut) saved as SVG. The formula engine is our own, so nothing leaves your computer to calculate.

Document: Markdown with a live preview, saved as Word, OpenDocument, PDF, a web page or Markdown. It never adds numbers you did not give it. PDF page tools split, keep, reorder (10-1 reverses) or join.

Plan: a room list becomes an isometric floor plan with seat counts, a legend and yellow highlights for changes, in four looks. Save as SVG or PNG at HD, 4K or 8K.

Your own files: Document opens Word, OpenDocument, web pages, RTF, Markdown and text. Sheet opens Excel, OpenDocument sheets and CSV. Files are converted on your computer. Work saves as a .vkstudio file you keep.

Animate
27 motions. The model writes short lines, one per element (canvas, text, count, shape, bars), about 1.5 times faster on the model call (6.3 s against 9.5 s), and a whole run on the server dropped from 40 s to 16 s. Code repairs what it can, and the model is asked again once for what code cannot fix. The result is plain HTML and CSS keyframes, so it needs no script to play.

Save Lottie exports live text and shapes, fades, slides, pops, growing bars and drawn lines. A few effects (blur, type-on, wipe, iris) fade in instead. OWNER: add the result of playing a file in a real player.

Save PowerPoint (editable text) writes real text boxes and tables (a:tbl). Checked in LibreOffice Impress. OWNER: add the Microsoft PowerPoint check. Known limit: text sits in plain boxes, not placeholders, so Outline view shows slides as empty.

Memory inbox
The tool the assistant calls to remember now writes to a small file on your computer. Only a Keep stores it, and what you edited is what is stored. Anything shaped like a password, key or token is refused before it reaches the inbox, and it holds 50 items. A setting turns the inbox off. Known limit: this covers what the assistant chooses to remember. Your own saves and Desk’s notes after a chat work as before.

Library skills
VEKTOR ships 44 skills (our 24 and 20 everyday ones). The other 157, adapted from open-source knowledge-work plugins, live in the Library: install one (about 7 KB) or all (1.1 MB). Each download is checked against its checksum, and an existing skill is never replaced. The Desk suggestion needs a word of your request in the skill’s name plus a fixed score. On seven test requests it picked the right skill for three and offered nothing for the four with no match.

Status-bar chip
It reads four things: graphics memory and heat, CPU load, free memory, and whether the local model is on the GPU or spilled to the CPU. A state changes only after two readings in a row agree. On recorded traces (steady, a spike, a game launching and quitting) it changed once per real change. It only watches. GPU readings need an NVIDIA card; without one it checks CPU and memory and says so.

Try it today
Update VEKTOR, open Studio > Image > Generate, and press Make it. For the add-on, press Install in Generate or Config > Local models. The changelog lists every change in 2.0.6, and the guide has new chapters for the bar, spreadsheets, documents, floor plans and motion graphics.

Improvements in 2.0.6
Studio and pictures

Use in Design: colours found in a picture in Animate become your Design colours in one click.

Quick starts and Recent: five starting points, and your last ten pictures one click away.

Upscale: 4x with an upscale model, or a built-in sharpen.
Design workspace for decks: layers, an assets manager, snap guides, presenter view, live charts and Morph transitions. An AI chat makes checked edits only.

Templates by business type: live in the content library, so the install stays small.

Storyboard page for Motion: one page with a player and a card per scene. It opens offline. vektor motion makes it from the terminal.

Timeline: curve handles, path tension, up to five masks on a clip.

Looping and voice: Bring the picture to life repeats exactly every 10 seconds. Each scene grows to hold its spoken line.

Plain advice on your video plan: too many words for a scene, hard-to-read colours.
Resume looks: six layouts with an accent colour.
Sharper pictures: Art saves at HD, 4K or 8K; new contrast-checked palettes.

Tables in decks: add a table slide, paste a CSV.

Moving backgrounds in Motion: particles, grid, waves or aurora.

Instant edits in Animate and Design: slower, faster, bigger text, move or rename a slide. No model, no wait.

Safety, memory and speed

Deleting always asks, even in Full access.

Steady memory in a chat: pinned memories are read once per chat window.
Plain requests on a local model answer in one go. One code edit went from 41 s to about 5 to 11 s.

Your message goes first: a quick question after an answer came back in about 2 s instead of 5 to 8.

Less re-reading each turn: prompt reading dropped from about 1.7 s to about 0.4 s where it applies.

A cloud model that does not answer is skipped after 12 s, not 30.
Optional faster engine for edits: set localEngine to llamacpp for about twice the speed on local code edits.

Smaller download: 36.4 to 31.0 MB.

Hidden console windows: git, python and helper launches no longer flash a window.

Agent and Coder

Coder inside Desk: an Auto, Chat or Code switch.
Plan first, See what changed, Work log, Background runs, Task list and questions.

Teams of helpers: up to four helpers on separate parts, after you approve. vektor agent.

Lessons, rules (AGENTS.md), commands (/review, /bugfix, /test and more), hooks.

Find definitions and uses across JavaScript, TypeScript, Python, Go, Rust, Java and more.

Scripts and batches: --json, --stream-json, --stdin, --until "npm test", --batch list.txt.

One shared service: vektor daemon start, plus a VS Code extension where every write asks first.

Desk, Vek and memory

Ask Desk about your email (read only), research agent (8 steps, 5 pages, 30 runs a day), ask Desk for a video (after you confirm the cost).

Sync, Import and Review (Config): two of your computers in step with no cloud; import ChatGPT, Claude or markdown exports with a preview.

Vek can use your MCP tools and asks before anything that changes something.

Routines are followed step by step. Models: Claude Fable 5.2 is in the Claude and OpenRouter lists.

Privacy and installs

Google sign-in asks for one permission, to send email.
Secret Hygiene skill checks for leaked secrets and tells you what to rotate first.

Installs are checked on every Node version (24.1, 24.22 and 26.13).
Backups hold recent notes; Library ratings and highlights are saved in your VEKTOR folder.

Fixed

“Failed to fetch” after a successful picture from a free provider.
p5.js sketches in Animate run again.

The Plan tab’s 3D floor plan draws a tall room in front of the desks behind it.

Coder reads large files a section at a time.

Card pictures that were slow now appear, and a failed picture search retries after 90 seconds.

3D and animation pages in the Sandbox run with no internet.

Gmail mail shows sender, subject and date.

A run that reaches its step limit offers Continue.

The Desk box no longer covers the bottom of Studio.
Guide: new chapters T2–47 to T2–51.

Meoow

Run your own race, keep your data sovereign.
Version 2.0.6 is live in the downloads page: https:www.vekmem.com

Full list of changes: https://vektormemory.com/docs/changelog

Novel Technology · Privacy Enhanced · Data Sovereignty · Vektor Memory

600+ models BYOK · 60+ MCP tools · 16 Model Providers

Top comments (0)