Novel Technology · Privacy Enhanced · Data Sovereignty · Vektor Memory
600+ models BYOK · 60+ MCP tools · Ollama Local & Cloud · 16 Providers
Studio Panel
We combined our motion and video timeline into one editing tool.
We are now more than just a memory layer; with Studio, you can now create clips, pictures, text and sound on tracks, trim and split, add captions and a voiceover, preview, and export. Compound clips and linked audio and video work, with undo and redo.
There are 540 ready-made templates across business, events and social formats.
The shelf is a new feature strip under the Studio tabs. Drop a picture, video or sound on it and send it to Timeline, Motion, Dither, Animate, Design, Transcribe or Media Factory without saving and re-adding it. A Back button returns to the tab you came from, up to ten steps.
Shorts takes a long talk or video and finds the strongest 30 to 60 second moments, titles them, and builds a vertical project with captions you can edit and export one clip at a time. It centre-crops. Face tracking is not in yet.
Auto in Motion starts from one line, such as “a 20 second launch teaser for our new product”. It works in stages you can watch: it reads the idea (length, shape, tempo and look, with the reasons written out), writes the plan, sets the tempo and look, checks the quality, and hands it over ready to play.
Studio also learns your checks, you teach a rule once, for example, a scene with more than 3 points, and it runs with Check quality on every video after. Pin notes on a scene or the whole video and press Revise. The assistant proposes edits for every note and you tick the ones you want.
Coder edit refinement that cannot quietly break a file
The agent has a new edit_file tool. You give it the exact old text and the new text, and the old text must match once. A miss tells the agent how to fix the request.
An edit that would leave the file with a syntax error is refused, with the error and the lines shown, and the file stays as it was. Rewriting an existing file of more than 50 lines in one go is refused too, unless you say you mean it. Every successful write ends with a quick syntax check, and in Coder the project’s own tests run after each write.
In the Coder editor the same idea reaches you. The Fix button on a problem asks the model for a corrected version and shows it as red and green lines. Accept puts it in the file as one edit that Undo can reverse, then checks again.
Reject drops it. If the file changed while you were deciding, Accept tells you instead of overwriting.
The agent also watches itself for wasted motion. It is nudged, then stopped, when it makes copies of files such as app_fixed.js instead of editing, rewrites a file much smaller, rewrites the same file three times in a row, or hits the same failure again and again.
If it calls the same tool with the same arguments and gets the same result, it is warned on the third time and stopped cleanly on the fifth, with progress saved so you can continue.
Routing that learns your accounts
Earlier routing assumed the accounts we use. This release learns yours.
When a provider limits or refuses a model, VEKTOR tries the other free models that provider lists: OpenRouter’s :free models, NVIDIA's free set, Cloudflare models that are not plan-only, and Ollama Cloud's list. It remembers what worked, what needs a plan and what was withdrawn, for you only.
If two different models hit a limit together, it concludes the limit belongs to the account and moves on. The provider order puts free providers first, and paid ones join only when paid fallback is on.
Keys are handled the same way. When a key hits a limit or is refused, it rests for ten minutes and the next one is tried. That covers numbered keys entries in the Passport.
Two measuring tools sit behind this. Find best models reads each connected provider’s current list, ranks it by size and recency, and checks the biggest few with a plain answer, a tool call with the right numbers, and a short function it must get right. It keeps the best model that passed, tries the next if that one stops working, and rechecks weekly.
The provider probe sends one tiny question and one tool call to each provider and records uptime and speed, never a key. A fast, reliable provider goes first. One that is out of usage goes to the back.
When a provider is slow, Desk hedges. If it has not started answering after a couple of seconds, Desk also asks the next one and uses whichever answers first, with a short note. It never uses a local model or a paid provider you have not allowed, and deskHedge set to false turns it off.
The Model keys panel puts this in words. A banner names any provider that has run out or had a key refused. Each row says what is wrong (out of usage, key refused, needs a plan) and links to the account page. A provider with several keys shows “6 of 10 keys work”, and Test my keys asks each key one tiny question.
We rebuilt our coder module to plan in steps
Coder writes a plan you can read and change, covering the files, the steps and how it will be tested. It makes a project folder and has the agent write the code. Then it checks the result for real: it runs the tests it wrote, checks every script and page, and flags anything the plan promised but never wrote.
Failures go back to the agent for up to three fix rounds. A second reader reviews the code against the plan. The finished app opens in Preview, and a checklist shows each stage.
You can choose a stronger model for the plan, the fixes and the review while the Coder model does the bulk writing. The run writes inside its own project folder without asking for each file, because you approved the plan.
If a model cannot use the file tools, Coder writes the files straight from the model’s answer. If the model cannot answer at all, it says so plainly.
Preview shows the workspace’s index.html in a sandboxed frame that cannot reach the app. Projects that need a dev server, such as React or Vite, tell you to start it and type its address.
Knowing your memory is healthy
Three new tools in Verify answer different questions for users.
The memory value report answers what your memory did. It is a one-page, printable summary: memories kept and written, what was recalled, how the memory gate kept things clean (new, merged, duplicates blocked), your busiest agents and spaces, and how searches went. It uses counts only, never your memory text, and nothing leaves your machine.
The memory quality check answers whether it is in good shape. It gives a score out of 100 and a plain list of what to fix. It asks for 24 of your own memories using a few of their words and counts how many come back.
It also looks for repeats, faded memories nobody uses, newer facts that contradict older ones, and untidy entries. It reads your memory and never changes it.
The clean-up tools act on what the check finds by reviewing repeated memories lists duplicates, keeps the oldest copy and lets you untick any before the rest are removed. A backup is made first and pinned memories are never touched.
Entity resolution finds names written several ways (plurals, capitals, spacing), merges them under one name, and lets you undo a merge or split a spelling back out, again with a backup first and no change to memory text. Decision lineage shows where a memory came from, why it was saved, which versions came before and after it, and which searches it went on to answer.
After an embedding-model upgrade, searches could return the same few memories whatever you asked. Searches and new memories now use the upgraded model, and the quality check can rebuild search numbers that no longer match their text.
You can ask Desk for an animation, a simulation, or a 3D scene
It now tells the model what it can draw, then opens the model’s code in the Sandbox with a live preview. three.js and p5.js are built in, so it runs with no internet. Pages that load them from the web or use an import map are fixed up automatically.
Once a program is open in the Sandbox code box, every model, small local ones too, is shown it and told to edit that program. Ask for “add rings” or “make it faster” and the model replies with the whole updated program.
Desk puts it into the code box, the preview runs it, and an Undo button sits under the reply. A snippet or a plain question about the code is never applied.
As an example, asking about the solar system gives you a moving model under the answer: drag to turn and tilt, scroll to zoom, click a planet for facts. Change review applies the same visibility to edits. When the agent changes files, you see the lines that changed in red and green and can undo any edit.
Several things that used to load from other sites now ship within the app:
To keep the app security and privacy controlled, we now host and load our own fonts, charts, the p5 drawing library, and the reCAPTCHA script. The package-safety lookup for skills runs only if you turn it on, and the anonymous model lookup is gone.
Speech, search and safety models now download from VEKTOR’s own host. Every file is checked against a fixed fingerprint, and a private mirror can be set in the config. The games in the library open full screen in a locked frame with no network, so they cannot reach your files or your notes.
**
Long chats and slow models speeds improved**
Older turns in a long chat are now shrunk to a short stub instead of being left out, so a fact from early in the conversation can still be recalled. The same handling now covers local Ollama chats in Lightning mode, which used to keep only the last four turns.
Streaming no longer cuts an answer on a fixed 60-second clock. An answer is cut off only when the model goes quiet: 30 seconds in the middle, 90 before it starts. If a model stalls partway, the part you already have is kept with a note and a Continue button. If a cut-off answer appears earlier in a chat, the model is told, so it does not treat the half-finished text as complete.
What we measured
Numbers from the release checks and our private benchmark scripts, run on one desktop (Ryzen 5 5600X, 48 GB RAM, Windows). Where a number is a replay and not live traffic, it says so.
A slow provider no longer slows Desk
Hedging asks a second provider when the first has not started answering. We tuned it by replaying 40,000 turns drawn from 30 measured first-token times per model.
OpenRouter free 550B alone: the slowest 5 in 100 turns fell from 13.6 s to 2.2 s.
NVIDIA 550B with the same model on OpenRouter as backup: the slowest 1 in 100 fell from 11.7 s to about 3 s, and the mean dropped 20 to 25%.
The cost: an extra call on 6 to 23% of turns, and another provider answering on 3 to 6%.
Live check, 8 runs, NVIDIA 550B with Groq as backup: median first chunk 3568 ms with hedging off, 2366 ms on.
The limits are real. The replay draws independent samples, while real queues are correlated, and a duplicate call can worsen a rate limit. Hedging doubles load on slow turns only.
On a local 9B model, median of 6 runs, time to the first streamed text.
With tools off Desk adds nothing measurable. The plain lane adds about 190 ms for the profile and memory lookup. Tool definitions add about 280 ms of prompt reading on a small model. Twelve turns of history add 5 to 25 ms. So a slow answer is almost never Desk. It is the model or the provider queue.
A trivial question, no Desk, median of 3 runs, time to a full one-sentence answer. Single runs spiked to 2 to 10 s on several of these. That variance is what hedging removes.
The guard runs per tool call, not per chat message. Median cost: 0.14 ms before a tool runs, 0.09 ms on a 500 character result, 1.1 ms on 10,000 characters, 10.4 ms on 100,000 characters, and 0.13 ms on a result carrying an injection attempt.
Other speed changes
Connectors used to take about 3 seconds to open. It now shows its last answer at once and refreshes behind.
Transparency draws 60 memories at a time instead of 200, about a third of the page elements.
A stalled stream is cut after 30 quiet seconds mid-answer or 90 before it starts, not on a fixed 60-second clock.
A provider whose keys are all out of usage is skipped instead of waited for.
Full regression tests pass on every release
The release gate ran 313 checks with none failed, covering unit, tool, security and integrity suites.
The live tool server passed 22 of 22 profile checks and the memory tools 19 of 19.
The extension manifest lists 71 tools and matches the source.
V2.0.3 — Changelog additions:
The agent and Coder
A run now survives a bad moment. Short rate limits are waited out, dropped connections and timeouts are retried up to three times, and progress is saved after every step.
If VEKTOR closes or crashes mid-run, the Agent panel offers to resume where it stopped.
Edits are safer. The agent replaces exact text, a change that would break the file’s syntax is refused, and rewriting a long file in one go is refused.
It notices when it is busy but going nowhere, nudges itself, then stops cleanly.
Build an app: describe an idea, get a plan you can edit, a project folder, written code, real checks, up to three fix rounds and a second reader, then a Preview.
A step limit is no longer the end. Continue resumes the saved run with 40 more steps.
Models and keys
Free providers go first. Paid ones join only when you allow paid fallback.
Routing learns your accounts: it tries the other free models a provider lists and remembers what worked.
A spare key takes over for ten minutes when one hits a limit or is refused.
Find best models checks the biggest models each provider offers and keeps the best one that passed.
If a provider has not started answering after a couple of seconds, Desk asks the next one too and uses whichever answers first.
The Model keys panel says in words what is wrong with an account and links to its page.
Studio
Timeline and Motion are one tool, with 540 ready-made templates.
The shelf moves a picture, video or sound between tabs without saving it first.
Shorts finds the strongest 30 to 60 second moments in a long video and builds vertical clips.
Auto turns one line of description into a finished video, in stages you can watch.
Studio learns your quality checks and revises a video from the notes you pin.
Memory
A memory value report gives you counts of what your memory did. It never uses your memory text.
A quality check scores your memory out of 100 and lists what to fix.
You can review repeated memories, merge names written several ways, and trace where a memory came from.
Separate spaces keep two clients from seeing each other’s memories.
Desk and everyday use
Desk opens animations and 3D in the Sandbox with a live preview that works offline.
Flashcards are real cards: flip, slide, mark Knew it or Again, then study only the misses.
Change review shows what the agent edited as red and green lines, with undo.
You can type while VEKTOR answers, and pick a permission mode for the chat.
Long chats keep early facts, and a stalled answer keeps the part you already have.
Fewer outside connections: fonts, charts and the drawing library now ship with the app.
**
**
600+ models BYOK · 60+ MCP tools · Ollama Local & Cloud · 16 Providers
Version 2.0.3 is live in the downloads page: https:www.vekmem.com
Full list of changes: https://vektormemory.com/docs/changelog#v203
Novel Technology · Privacy Enhanced · Data Sovereignty · Vektor Memory
Video Production
LLM
Superintelligence
Agentic Ai
Top comments (0)