Another gigantic update. We have now added Hacker News books and video recommendations into the library.
We added 1 million trial tokens that are included on all new trial signups to test out the systems via credit card.
We now include model GLM-5.3-Flash with 40 million tokens per month, per user on the $9 per month paid plans, as newer, faster models are released, we will upgrade to them. You can still use the free model provider options and Ollama local GPU as well, it’s your choice.
We upgraded to Node v26 engine for the latest security updates with automatic fallback to your own Node version and the Vektor upgrade node once ready.
We also upgraded to Better-sqlite3 v13: one binary for Node 22, 24 and 26.
We also upgraded to the latest ONNX version is now 1.30.0.
We added in procedural memory nodes to make a 5-layer graph: entity, causal, temporal, semantic, and procedural.
From the 9th of November, Ollama Cloud becomes a pay-as-you-go provider, as per their notice, we have updated this in our models section.
Books and videos from Hacker News that people recommended
Hacker News comments are one of the best reading lists on the internet, and searching them can be difficult. We combined both into our own list from the public comments: about 9,100 books and 27,000 videos, each with how many different people recommended it and when it was first mentioned.
It lives in the Library under HN SHELVES.
You can pick books or videos, view in Open Library (Today’s shelf, Top all-time, Rising, Recently recommended, Newest or Oldest first mention, Title A-Z), and filter by publication year, first-mentioned year, topic and features.
Click a card and the Explain panel tells you what the book is about and who it is for via an instant LLM response, using the model you already chose.
Pictures are off until you switch them on. When they are on, covers load only for the cards you scroll past, are shrunk to about 4 KB on your computer and are kept in a local cache.
Nothing is stored on our servers. If a book has no cover, or the source sends a blank one, you get a Bauhaus drawing made from the title, so the same book always looks the same.
We added procedural memory for how to do things
Facts are only half of what you remember. The other half is how: the steps that finally got a release out the door, the order you check a spreadsheet, the way you answer a hard email.
Procedural memory keeps those as routines. A routine has a short name, when to use it, numbered steps and what done looks like. A new one starts as a draft. After two successes it becomes active, and after three failures in a row it is retired. Secrets are cut out of every field before saving.
When a question fits an active routine, the steps are added at the end of the prompt, where a small model’s limited window keeps them.
You can write one by hand, paste what worked and let the model draft it, or turn a good routine into a reusable skill. The graph has a Procedural layer that links each routine to the memories it came from, and agents get five tools to find and save them.
Does it help? We ran a small test: ten simulated data tasks, each with specifics only the routine could supply, three runs each, on two models.
With no routine: 0 of 30 correct on both models.
With the right routine found by the real retrieval code: 22 of 30 on both (73%).
With a routine learned from a short account of a run that worked: 21 of 30 and 17 of 30.
With a routine for the wrong task: 0 of 30, the same as nothing.
The tasks are simulated, the scoring is literal, and the intervals are wide, so read this as “the idea works and the plumbing is right”, not as a benchmark to quote, as we will test this again with full benchmarks once fine-tuned next week with more data.
A model that comes with a small quota of 40 million tokens
Since the provider free tiers keep shifting and degrading, we added a provider called VEKTOR Included. Pick it in the model list and you chat on GLM-5.3-Flash.
The allowance: 1 million tokens for the trial week, then 20 million a month. When that is used up, a Reset Quota button gives you another 20 million once per month, so a total of 40 million per user.
Where you see it: a bar under the context window shows what is left and when it resets. Hover the card for the rules.
Privacy: requests pass through our server to reach the model. The server keeps one number per licence, a token count, and nothing you wrote. The provider is labelled in the picker and is never used as a hidden fallback.
Speed: the first real word of an answer arrives in about 1.3 to 2.4 seconds in our live tests. The model thinks first and shows the answer after.
We will try our best to increase quotas as newer, more efficient models are released and inference costs continue to decline. We can appreciate users want larger, higher-quality models and more tokens included.
Faster answers when the model is in the cloud
A question in Desk used to cost two model calls. With tools on, Desk first asked the model to decide on tools, threw away the answer it got back when no tool was needed, and asked again for the final text.
Now the first reply is used when it is the answer, and every round streams, so the first words show up as they are written. A question makes one call instead of two.
We also checked the NVIDIA free endpoint model by model. The default we shipped, gpt-oss-20b, timed out at 90 seconds. nemotron-3-nano gave a real answer in 1.1 to 1.3 seconds. The default is now nemotron-3-nano.
Studio: more of what editors expect
The Studio timeline had 20 effects, 8 transitions, captions, voiceover and a shorts finder added. The new one is Chroma key: it makes a green or blue screen transparent, with controls for how soft the edge is and how much of the green glow to pull out of hair and edges.
We also compared the timeline with other tools, feature by feature, and wrote down a plan to upgrade these over time: the first four are now one-click buttons in the timeline toolbar.
Auto cut removes quiet stretches and closes the gaps (an 8 second test clip with 2 seconds of silence came out 1.8 seconds shorter), Beats puts a marker on every beat, Duck lowers music under a voice clip, and Level sets a clip’s volume from its loudness. Each is one Undo.
Editing by deleting words in the transcript, noise clean-up and subject-following reframe follow. A permissively licensed open source suite turned out to describe most of these methods, which saved us from guessing.
Everything else in 2.0.4
New
HN SHELVES in the Library: about 9,100 books and 27,000 videos with recommend counts, filters, optional local pictures and Bauhaus covers.
Procedural memory: routines, the Procedural graph layer, five agent tools, graduate a routine to a skill.
VEKTOR Included: a model with your licence, a usage bar, a monthly Reset Quota button.
Chroma key in the Studio timeline.
One-click Studio edits: Auto cut, Beats, Duck, Level.
Answer guard: Desk answers end with a short list of rules, and a check for links nobody supplied. In a test of 14 invented questions run three times on two models, answers to requests for links nobody had supplied were clean 6 of 18 times without the rules and 18 of 18 with them on one model, and 4 of 18 against 17 of 18 on the other.
Faster
One model call instead of two for plain Desk questions on cloud models; answers stream as they are written.
NVIDIA default is now nemotron-3-nano. Latency shows in seconds in desk
A managed Node 26 engine with automatic fallback to your own Node, and vektor upgrade-node.
better-sqlite3 13: one binary for Node 22, 24 and 26.
Providers and cost
Ollama Cloud: all 24 models listed with prices, a PAID tag from 9 November 2026, and not used as a silent fallback after that day.
Download size: the voice code no longer installs a second copy of the speech runtime. The install is about 300 MB smaller and the unsupported-package warning is gone. Voices download once from our own mirror.
Look and feel
Quality editorial diagrams with a dark skin, and a bundled diagram-design skill.
Vek answers in its own voice: no forced one-sentence replies, no stock openers, and a short note of what you were working on lately.
Config: balanced two-column layout with no gaps, a half-width mascot tile, runtime checks for Node, ONNX and the database in About.
Explain on Hacker News cards, a round send arrow in Desk that becomes a stop button while it works.
Upgrade now:
Update with npm install -g ./vektor-slipstream-2.0.4.tgz and open the app.
Library, then HN SHELVES, is the quickest tour path.
Procedural memory is in Config, with an Add 5 examples button if you want to see the graph layer fill in.
The VEKTOR included model quota is in the model list and needs the licence you already have.
Run your own race, keep your data sovereign.
Version 2.0.4 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)