Apologies for the mind dump of updates below. We’ve been so busy building that I haven’t been able to keep up with the changes. It seems every day a new LLM or a new feature is released, so it’s hard to keep track.
We’ve put a real focus on improving our UX and fine-tuning the LLMs’ outputs. This is a core challenge, as every LLM gives different answers, and getting fine-tuning, guardrails, and consistent prompt accuracy right is mind-bogglingly hard. It’s very much like a finger in a dam: fix one thing and another issue arises. Providing quality outputs across 600+ LLMs is not easy to perfect in a harness.
We left personalization until last, with SOUL.md integration for a small bot that learns about the user and offers quips and advice. It’s still expanding, learning from the memories in the system itself, and becoming more aware every day, but the technology has come far enough that it’s relatively secure and not too annoying. Don’t like it; turn it off in settings.
We’re in two minds about whether it’s worth adding payments and Muse, OpenClaw or Hermes-level functions and still weighing whether that’s worth the hassle and the issues it causes, as rogue bots with payment portals and your credit card are dangerous without hitl; instead, we are staying more business-tools focused for now.
Depends on user feedback as to what people want and utilize, will review and go slow till all the unknowns are safe and solved. Being local first, private and secure is always our main focus.
This is a two-in-one update. v1.9.6 made the multi-agent side of VEKTOR visible and hands-on: you can watch Collab agents write files and run commands live, and they now work with real code in the same terminals you see. It also added a tenth free model provider, offline text-to-speech and a Business Coach.
v1.9.7 is about your local model functions. It can now see and use the majority of VEKTOR panels. Before this release, a small Ollama model would tell you it had “no access to a library system” while sitting on 632 indexed files. That’s now fixed, and as more models are entered, we will revisit often as LLM’s update.
There’s also a new personality in the sidebar. Vek, the little monitor robot, now talks through your local model, knows your name and your weather, and every so often shares a thought of its own. It never recites facts at you. Our motto for it is the one we use for the whole product: thinking, not retrieving.
Vek may get more features depending on feedback.
None of your data is stored by us. No ads, no telemetry, no embedding costs, no data caps, and nothing of yours is used for training. The mascot’s voice and thoughts run on your local model only, so they cost nothing and never leave your machine.
VEKTOR v1.9.6: What’s New
Collab you can watch
See Collab agents work live. File writes and terminal commands now show up in the interface the moment they happen, and the final answer streams in word by word. Before, you saw nothing until the whole task finished.
Collab agents work with real code. Worker agents can read, write and run code, and they drive the same terminal panes you see on screen, not a hidden shell. Risky actions, like deleting a file or running a destructive command, wait for your approval.
Quality checks are a vote, not one opinion. When a result is uncertain, up to 3 models check it instead of trusting a single verdict. Clear-cut results still resolve quickly and cheaply.
Collab recovers instead of failing. A run could fail completely over one bad provider key, even after every agent had finished and passed its checks. The final merge now tries each configured provider in turn. An old, discarded attempt from a retried agent also no longer leaks into the final answer.
New providers, voices and tools
Ollama Cloud, a tenth free provider. Free hosted models (gpt-oss 120b and 20b, gemma4, nemotron-3) are now a provider of their own, separate from local Ollama. No GPU needed.
Offline text-to-speech with Kokoro. A local voice that needs no API key and runs fully offline after its first download. Hume and Cartesia joined the cloud voice options.
Every model gets real tools. Some models, including Ollama Cloud and local Ollama, claimed they had no tools even when tools were set up. They now get the same tools as every other model, and SSH tools work with every provider.
A real weather widget. The Health tab shows current conditions, a 5-day forecast and humidity, and it can detect your location automatically. Asking about the weather in chat now actually looks it up.
Business Coach
Three coaching modes inside Desk. COACH turns a brain-dump into a reflection and up to 3 priorities written straight to your Tasks board.
CRUCIBLE is a guided decision dialogue (5-Whys, pre-mortem, weighted trade-offs) that saves decisions with 30 and 90 day review reminders.
CADENCE is a weekly habit and decision check-in.
Small fixes
Saving an API key in Config now shows it as saved straight away, with no page refresh.
The copy button on memory cards has a visible border again.
Desk has a persistent file-tree panel, matching the other side panels.
VEKTOR v1.9.7: What’s New
Local models that use their tools
Local models now reach your library, memory, config and Faraday. Desk used to send small local models all 81 tools at once, with the library tool at position 65. The model gave up and answered from guesswork. Now a local model gets a focused shortlist: about 12 core VEKTOR tools plus whatever matches the words in your question, capped at 20. Hosted models still get the full set.
Tool choice is steadier on local models. At Desk’s default temperature of 0.7, the model used the right tool in 2 of 3 identical runs. The first tool-choosing round now runs at 0.2, which took it to 3 of 3.
Desk knows its own app. Ask it to list the 10 Desk modes, the 15 nav panels or where your skills live, and it answers correctly. The map is generated from the real interface file, so it can’t drift from what you see on screen. It stopped inventing documents like a “skill_files.md” that never existed.
Browse your book library. “List a few books in my library” used to run a keyword search for the words “list” and “books” and come back empty. Now a browse request returns a sample of real titles plus the total count.
Desk as a code workspace
Desk can build straight into the Sandbox. Ask for a landing page, a component or an SVG, and the model opens it in the Sandbox editor next to the chat with a live preview. Under the hood a new sandbox_open tool streams a sandbox_load event that the interface picks up.
Files the model writes open in the Files panel. A successful file write now sends a file_written event, which opens the file live in the Files panel. It’s the same path Collab workers already used.
More code tools in Desk. Desk gets the Agent tab’s code search, grep, semantic code search, blast-radius and repo-architecture tools. They run through the same approval checks the Agent tab uses.
The mascot
It has a voice. After each answer, Vek reacts with a short line from your local model. Click it and you get one useful nudge grounded in your real memories. A hard rule stops it from mentioning anything that isn’t in the memories it was given.
It has thoughts of its own. Every 7 to 14 minutes it may share one small original thought, sparked by the time of day, your weather or what you’re working on. It only speaks while you’ve been active in the last 20 minutes and the app is on screen, never mid-answer.
Robot jokes and sci-fi nods. About a third of its thoughts are a self-aware robot joke or a nod to a film like Blade Runner or Star Wars. One real line from testing: “Fixed the panel, but did I just see the replicants among the files?” If it copies one of its example lines word for word, it regenerates once.
It has a SOUL.md. Vek’s personality lives in a plain markdown file at ~/.vektor/mascot/SOUL.md, which you can edit in Config. The six personalities (cool, funny, caring, helpful, intelligent, smartass) sit on top as a mood, so the tone shifts but the soul stays the same.
You can rename it. The default name is Vek. Change it in Config and the speech bubble label and the mascot’s own sense of its name both update.
It knows your name and your weather. Your name comes from your Profile, falling back to your system username. The weather comes from the location you set in the Health tab, cached for 20 minutes so a reaction never waits on the weather service.
A 2.5D look in your theme colours. The mascot is now a flat-shaded monitor robot built from three surface tones, so it follows all eight themes. It has a bigger screen face and body language that matches its expression: it hops, tilts, shrugs and wiggles.
Faces stay on the screen. Its little screen shows only short keyboard faces like ^^ and ><. Anything wordy goes to the speech bubble, and longer faces shrink to fit.
Don’t like it; turn it off in settings. With more customization to come in the future.
Voice
Listen starts almost instantly and reads the whole reply, via a local LLM. Replies used to go to text-to-speech as one request, capped at 4,000 characters. Now they’re spoken sentence by sentence: the first plays as soon as it’s ready while the next ones are prepared in the background. On a normal Windows machine with Kokoro running locally, a 357-character chunk took 1.4 seconds to prepare, far quicker than it takes to play.
Cleaner speech. Code blocks, links and markdown symbols are skipped, so it no longer reads out asterisks or source code. Hovering the Listen button warms up the local voice model, since loading the model was most of the delay.
4 voices added; if more voices are created, we will add them in later on.
Workflows
Tasks, Skills and Prompts in one place. Saved Collab tasks can be grouped into your own categories with filter tabs. Agent Skills are managed on the same page, and a new Prompt Gallery sits alongside them.
A saved workflow uses its preferred model. The preferred model on a workflow used to be a label only. Now it’s the model that actually runs.
Type / for a command palette. Typing / in the chat box opens a searchable list of every mode, including LIGHTNING, COLLAB and COUNCIL.
Updated a lot of workflows and templates
Interface polish
Softer, deeper, more consistent. Borders dropped from 26 to 40 percent opacity down to about 9 to 15 percent, and buttons, cards and the composer got a subtle top highlight and shadow. There’s one filled accent button per view, with squared-off corners and a keyboard focus ring everywhere.
Cards that line up. Every workflow card now uses the same fixed layout, so buttons, tags and descriptions sit at the same height across the whole grid.
Readable text sizes. 512 font sizes below 10 pixels, some as small as 7, were raised so nothing sits below 9. Buttons now inherit the text size around them instead of the browser’s default.
Fixes
The code-preview sandbox no longer goes blank white after you close and reopen it. Closing it now unloads the page, so every reopen starts fresh.
Notes saved from AI output no longer show a raw code dump as their title.
The Files panel tree now fills the panel when no file is open. It used to stop at 240 pixels.
The Sandbox, Terminal, Files and Flowire buttons no longer turn dark after you close their panel.
BUILD mode no longer produces pages with literal \n text when a smaller model mis-escapes its output.
Flowire-assisted edits are now checked for real syntax and type errors right after they’re applied.
A Collab run’s answer no longer fails to render silently.
Upgrading
Drop-in from any prior version, same as always, and v1.9.7 includes everything from v1.9.6. Your memory database stays untouched.
npm install -g ./vektor-slipstream-1.9.7.tgz
Full changelog with everything not covered here is at vektormemory.com/docs/changelog. Questions or feedback, the forum’s the fastest way to reach us.
VEKTOR Memory builds local-first persistent memory infrastructure for AI agents. Documentation and downloads at vektormemory.com.
LLM
Agentic Ai
Agentic Rag
Chatbots
Top comments (0)