Agentic memory you can trust via any method you decide to use it.
Custom code image
An agent with memory is only as good as its weakest read path. If a note can silently truncate, if a model switch can silently fall back to something you didn’t choose, the system isn’t really persistent. It’s just quiet about the parts where it fails.
v1.9.0 is built around a single premise: consistency is a feature, not a maintenance chore. This release makes VEKTOR behave the same way whether you’re in Desk, the Agent tab, JOT, or the terminal. Same provider logic guarantee and connector security.
The provider waterfall, unified
Every interface in VEKTOR now follows one rule: try the model you selected first, then cascade through every other provider you’ve configured, and treat a local Ollama model as the genuine last resort, not the accidental default.
If you like, you can set your Ollama LLM first and use 100% local and private inference with an open-source model.
The Agent tab’s chat and the CLI each carry their own provider logic, built at different times, and both had quietly fallen behind. Select Cerebras in the Agent tab and it would drop straight to Ollama, because that code path only recognized four providers. Select Gemini or Mistral from the CLI and the same thing happened for a different reason: both were listed as valid choices with no implementation behind them at all.
We rebuilt this as one shared chain, used by all four surfaces. In practice, that means a request now survives real failure. A live test after the fix routed through Claude, then OpenAI, then landed on Groq, each failure logged and handled in sequence rather than surfaced as a dead end. The model you picked is the model that runs, and if it can’t, you find out why instead of getting an answer from something else with no explanation.
Desk brief updated
Updated provider list
Supported across Desk, Agent tab, JOT, and CLI (same waterfall logic everywhere):
Claude (Anthropic)
OpenAI
Groq
Gemini (Google) — newly added to CLI this release
Cerebras — newly added across GUI, Agent tab, JOT, and CLI this release
Mistral — newly added to CLI this release
xAI (Grok)
OpenRouter
Ollama — final local fallback, unless chosen as primary
Models:
OpenAI
Astra 6
Cerebras
qwen-3.8-27b
gemma-4-31b
gpt-oss-120b
Gemini
Removed gemini-2.5-flash-lite (Google deprecated it for new users)
Default now points to gemini-3.5-flash-lite
Groq
qwen3.8–27b added to the lineup
Twelve connectors, one security layer
VEKTOR now ships 12 live connectors: GitHub, Slack, GitLab, Vercel, HuggingFace, Jira, Linear, Sentry, and Notion, fully tested and running today. Gmail, Google Drive, and Microsoft Teams ship OAuth-gated on your own app credentials, not a shared client key, because those three touch a real inbox and a real file system.
Connector panel
Every one of them sits behind Faraday, the security gate that screens MCP tool calls before they reach memory.
This release adds OWASP LLM Top 10 and MITRE ATLAS classification to every event Faraday logs, and a behavioral anomaly baseline that flags a tool call whose frequency breaks from its own history. Faraday’s signature self-test runs 23 known attack patterns against 23 benign controls. Zero false positives, both directions.
The agent engine, restored to full strength
This release rebuilds and improves all 53 functions in the autonomous agent core: semantic code search, subagent delegation, lint and test tooling, code review, self-evaluation, commit message generation, and the hook system that ties them together.
Built to catch its own mistakes
A golden-recall suite runs fixed queries against fixed seed memories with known expected results, built specifically to catch the failure mode where a system still answers, just not correctly. A smoke test sweeps roughly 240 files on every install. Both now gate every commit before it lands, not after a user finds the gap.
The CLI regression suite runs 41 checks across every surface touched this release, making sure all passed the quality code inspection.
What this means going forward
Persistent memory is only as trustworthy as its least-tested edge. This release closes several of those edges at once: a note that no longer loses its own content, a model choice that no longer gets silently overridden, and a connector layer with the same security guarantee no matter which one you’re using.
The current release builds on our privacy and security foundations rather than patching around it.
Why we build it this way
None of this works if the underlying architecture isn’t private by default.
VEKTOR computes embeddings in-process, on your device, with a local ONNX model. There is no cloud copy of your memory graph to log, breach, or subpoena, because there’s no cloud copy at all.
That’s not a policy we ask you to trust. It’s a property of the design: air-gapped by default, so surveillance is structurally impossible rather than contractually prohibited.
Your memory lives as a full SQLite file, on your own machine, with a real file path.
It’s never moved to mandatory cloud storage, never used to train a model, your data is never sold for ads, and it’s yours to copy, migrate, or walk away with at any time via our open-sourced DB migration tools.
We built VEKTOR this way because a memory layer that centralizes someone’s most private conversational history has no business asking for blind trust.
It should be something you can verify yourself.
Get v1.9.0
Read the full changelog → https://vektormemory.com/docs/changelog
Vector Database
Agentic Ai
Information Security
Ai Memory
Top comments (0)