{"slug": "vektor-2-0-8-studio-timeline-faraday-security-longmemeval-89-6-improvements", "title": "VEKTOR 2.0.8: Studio timeline, Faraday Security, LongMemEval 89.6% Improvements", "summary": "The VEKTOR project released version 2.0.8 of its memory and agent platform, adding MCP security checks in its Faraday tool that block tool-description rug pulls, cap sub-agent tool access, and re-screen memories saved while suspicious content was on screen. The release also reports a score of 89.6% on the public LongMemEval benchmark, up from 84%, attributed to broader search over saved history for questions about dates, counts and durations, alongside Studio timeline video-editing upgrades.", "body_md": "Today we’re excited to release VEKTOR v2.0.8. Major improvements to our security tool Faraday for MCP prompt injections and rug pulls. Studio timeline feature upgrades and LongmemEval improvements to reach 89.6% across the 6 categories below.\n\nDemo: [https://youtu.be/-mOfmFxyeX0](https://youtu.be/-mOfmFxyeX0)\n\n**LongMemEval Benchmark:**\n\nOn the public LongMemEval test (500 questions, the easier “oracle” version, graded by two open models, one run), the full run now scores 89.6% on average, up from 84%. The gain comes from searching more of what you have saved when a question is about your own history, such as “how many”, “how long” and “when did I”.\n\nThe idea behind the release: your files should become something you can ask questions of, and an agent should earn trust before it gets any.\n\n**LongMemEval Results**\n\nTry it here: [https://vektormemory.com/downloads](https://vektormemory.com/downloads)\n\n**Faraday MCP security tool:**\n\nFaraday now checks more of what your agents do. A helper agent cannot be handed more tools than the agent that started it, and a task steered by something it read on the web is held for your answer.\n\nMemories saved while suspicious content was on screen are checked again when they come back, so a poisoned note cannot quietly steer a later session. A tool that changes its description after you approve it (a rug pull) is still caught and blocked. Plans and READMEs written as lists are no longer mistaken for hidden instructions.\n\n**Studio video editor:**\n\nThe Studio timeline gains real text with a video editor’s controls, 18 animated text styles, one-click Smart cutout, motion tracking, enhance and AI upscale, and backgrounds that build themselves while you work.\n\nIn the Agent profile, Content reads the key details from each file (parties, dates, amounts, vendor, totals) and shows them as a table. Anything uncertain is marked, so you check only what needs a look. Then you ask questions of the table: “which are over 10,000?” and you see how the answer was worked out. Export to CSV whenever you like.\n\nEach detail is read once and kept with the file and the place it came from. A question runs over the table, not over a fresh guess, so the same question gives the same answer. Where a detail is missing, a form can collect it, and a Word, PDF or Markdown document is made from the details with the gaps marked instead of guessed.\n\n**How does an agent get only what I allow?**\n\nYou give an agent access to a space up to a level: public, internal, confidential or restricted. It receives the few passages it needs, each with where it came from, so answers are shorter and cheaper. A passage above its level never shows up for it at all, so there is nothing to talk it out of.\n\nYou also set what an agent may do. For example, only tidy up in one folder. Moving a file to the trash is always a request you approve, and it can be restored. People get roles (viewer, editor, owner) and the highest level of content each may see. Known limit: this controls what VEKTOR hands to an agent. It cannot control what a person copies out of their own screen.\n\n**Can I describe a process and get a workflow?**\n\nSay what should happen: “when a contract arrives, if the value is over $10,000 ask me to review it, otherwise archive it”. VEKTOR writes the draft in plain words. Try it on your files first. A trial changes nothing and sends nothing. Turn it on when you are happy, and it then runs by itself on new files.\n\n**How do I build an agent, and know it works?**\n\nBuild starts as a short chat: what should it take care of, how should it sound, what must it never do. It is mostly taps. You see what will be created, press Yes, and it is ready. Later you describe a change, see it line by line, and roll back any version.\n\nBefore you rely on it, run the simulations. They act as awkward customers (angry, vague, trying to make it break its rules) and check the answers. A change cannot go live until the checks pass or you choose to go ahead. When something fails, the fix is drafted for you to review. Known limit: simulations find the failures we thought to ask for. They are a floor, not a guarantee.\n\n**Improvements to the Studio timeline**\n\nTitles are real text. Font, size, weight, alignment, spacing, outline, shadow, a box behind it and seven presets, like a video editor.\n\n18 animated text styles (typewriter, fade up, glitch, karaoke and more) in Motion and the timeline.\n\nSmart cutout. Select a picture, press Smart cutout, click what to keep, right click what to lose. Even a busy background comes away cleanly.\n\nTrack motion. Drag a box around something moving and a title or sticker follows it. Two boxes follow turn and size. Four pin a picture to a moving sign.\n\nBackgrounds that build themselves. Auto backgrounds fills a track with a different look for each stretch, then swaps in Wan 2.2 pictures as your graphics card finishes them.\n\nEnhance and AI upscale. Clean and sharpen, brighten, punchy colour, 2x for clips. For pictures, 2x or 4x upscale.\n\n**How does the coding agent build a bigger app?**\n\nBuild keeps a feature list and progress notes inside the project, finishes one feature per run, runs the checks itself and saves each finished feature. If a build stops, it picks up where it left off. A feature that fails three times is set aside so the rest still gets built.\n\nWhen the build makes a web page, VEKTOR opens it in the Edge or Chrome you already have and fails the check on script errors, missing files or a blank page, with the error shown.\n\nThree helpers can look but not change your files: Explore reports where things are, Test runs the checks and says what failed, and Review lists bugs most important first. A project map lists each file with its functions, so the agent opens the right file the first time.\n\nA task handed to the coding agent works in its own copy and branch. You see what changed, then decide. A draft pull request needs your confirmation, and nothing is merged for you.\n\n**How do I stay in charge of what agents do?**\n\nGuardrails: choose Careful, Balanced or Hands-off, set quiet hours, and apply it to everyone, one agent or one routine. Deleting, paying, emailing outside, publishing and installing always ask first, whatever you pick.\n\nThe run timeline shows what started a run, which memories it used, each action and how it ended, with the slowest step named. An audit timeline shows every change by a person and every action by an agent, and customer details are never copied into it.\n\n**Further Improvements in v2.0.8**\n\nDocuments and agents\n\nKeep it protected. Move a space to a folder or storage service you choose. Files are locked with your passphrase before they leave this computer, and any change to a file is noticed.\n\nSee what customers talk about. Drop in conversations and get the topics with counts and a weekly briefing of what is rising.\n\nOne inbox for everything waiting on you. Team alerts, customer messages and pull requests arrive as drafts. Nothing is sent until you press Send and confirm.\n\nSmall jobs that run by themselves. Thirteen ready-made loops (a weekly wrap, a check after every merge) start switched off and begin with a dry run.\n\nPersonal touch for customers. Keep what you know about each customer.\n\nYour agent offers the right thing, or nothing, and opt-outs always win.\n\nOpportunity alert. Write what you found, the proof and the next step, and it waits in Drafts as a Slack or Teams message.\n\nThings to improve. Write what should get better, press Do it, then enter the number afterwards to see whether it helped.\n\nQuick first, advanced below. The Agent profile has sections with a search and one agent picker. The section bar no longer jumps when you switch.\n\nRoutines from your files. Turn a written procedure, transcript or recording into steps you check and approve, then into a skill.\n\nWhat should I do next? One press names the one thing worth doing now. At most three a day, and nothing in quiet hours.\n\nOutcomes. Type what you want to happen and get a short list of steps with dates. No model needed.\n\nGuardrails and run timeline. Choose how much an agent may do alone, and see each run step by step with the slowest step named.\n\n**Coding agent**\n\nThe Coder fixes by itself. A failed check gets one plain fix first, then three different fixes side by side in separate copies of your project, and the best one is kept.\n\nA Build ends with a short card. What was built, whether the checks pass, one question and one or two buttons.\n\nUpgrade now. A button next to the update notice in Config. Your data is copied first, and VEKTOR goes back to your version if anything fails.\n\n**Memory**\n\nMemory playground. Ask one question and see the answer without your memory and with it, side by side, with the memories used.\n\nMore accurate memory answers. Questions like “how many”, “how long” and “when did I” look through more of what you saved, and new memories keep what happened and for how long.\n\nImport into memory. List files or public web pages, check what each would do and what would leave your computer, then import.\n\nFive small tools in Config. Suggest edits, Decision log and three more.\n\n**Connections and setup**\n\nTen more connectors. Trello, ClickUp, monday.com, Pipedrive, Mailchimp, Toggl Track, Typeform, Webflow, Netlify and Fathom Analytics.\n\nConnect other tools. Add VEKTOR to Claude Desktop, Claude Code, Cursor, VS Code or Windsurf in one press, after a confirm and a backup.\n\nPlugins and saved prompts. One folder of skills and helpers per workflow, and prompts with blanks like {{client}}.\n\nSafer, easier skills. Saving a skill points out a missing description, project skills run only after you trust the project, and a loaded skill stays in force through long runs.\n\nMore Preferences. Contrast, popups that stay, your own code font and size, branch prefix, where working copies live, and keep the computer awake while an agent works.\n\nAdd-ons keep themselves current. Check for updates lists each add-on and repairs what is missing.\n\nModel lists. GPT-6.2 Astra and GPT-6.2 Sol through OpenAI and OpenRouter.\n\nThe Three things to try box goes away by itself once you send a first message.\n\n**Studio timeline**\n\nMatch look. Pick a clip and one whose colour you like, press once, and the first moves toward the second.\n\nCutout. Remove a plain background, such as a product on white, with one press.\n\nSlow zoom. A gentle zoom and drift turns a still picture into a moving shot.\n\nAn AI app can edit your timeline. Seven tools let Claude or any MCP app build and change a project.\n\nMake it better in Motion. One press adds transitions, animated text, a moving backdrop, captions, a closing line and a film look.\n\nCamera look. Pick a lens, light and camera move, and Wan 2.2 pictures and clips follow it.\n\nAudio to video. Turn a song or voice into bars, a waveform or a scrolling spectrum on its own track.\n\nAuto colour and Palette. Balance dark or tinted footage in one press, and click any of a clip’s five main colours to copy it.\n\nImage to video. Turn a picture into a short moving clip with Wan 2.2.\n\nWan clips look sharper. More steps and a cleaner encode by default, and Wan 2.2 14B makes clips from words alone.\n\nSnip has a pen and a highlighter. P and H are the shortcuts.\n\n**Studio plans, voice and publishing**\n\nFive caption looks. The bar, outlined text, a pill, big bold words, or word by word.\n\nRecord your own voice for a scene. It replaces the computer voice and the scene stretches to fit.\n\nHook ideas and headline ideas. Five opening lines or cover headlines from your material. A line with a number that is not in your material is dropped.\n\nChapters, a post description and alt text from the plan, with no model and no number invented.\n\nA teleprompter. Large scrolling text at 90 to 220 words a minute, with a mirror switch.\n\nDub a video in one press. Pick a video and a language, and the translated voice is laid under the picture.\n\nTranscribe tells speakers apart. Pick 2, 3 or 4 speakers and the text starts each change with “Speaker 1:”.\n\nCover pictures for every place you post, sized for YouTube, TikTok and Reels.\n\nRewrite text for a platform. A YouTube Short script, TikTok caption, LinkedIn post, X thread or short email. Every number, date and name is kept.\n\nPin notes on scenes, then Revise. “Say this more plainly” on a scene, and your model revises.\n\nStudio reviews its own video. Every scene on one sheet, with text cut off at an edge or an empty scene flagged.\n\nSnap cuts to the beat. Studio finds the tempo and moves each scene change onto the nearest beat.\n\nA waveform with markers under the preview, and sound checks for clipping, long silence and music as loud as the voice.\n\nAuto learns your usual style, offers ready-made starting points (Social short, Logo sting, Product demo, Explainer, Story), and can add stock footage from your own Pexels or Pixabay account.\n\nMore advice on pacing, and the Ask bar can add a scene (“add a closing scene that tells people to try it”).\n\nYour own brand font in a deck, a look inside any template before you use it, earlier plans kept so you can fill the boxes again, and Studio Art keeps its last 12 designs.\n\n**Faraday MCP Security tool (safety checks)**\n\nHelper agents stay inside their limits. A helper started by another agent cannot be given tools its sender does not have, long chains of handoffs are stopped, and a task that repeats an address or detail from outside content you never wrote is held for your answer. Approve it once and the same task goes through.\n\nThe Agent tab is covered. Writes, code runs and tools from connected apps are checked the same way as in Desk. A call that looks steered by something it read on the web is held with the reason, and what web and app tools bring back is checked on the way in.\n\nSaved memories are watched. A memory saved while suspicious content was on screen is remembered, and if it comes back later the session is checked again before anything is sent or changed.\n\nFewer false alarms. Plans, READMEs and notes written as headings and lists are no longer mistaken for hidden instructions, with the same attacks still caught.\n\nA plain-language report. What was held, approved, blocked and flagged in a session, as a short readable page.\n\nFaraday keeps checking on your own tools. A call with an email address or phone number could skip later checks. It no longer does.\n\nDesk and look and feel\n\nLonger, clearer answers. A reply that runs out of room carries on by itself, in short paragraphs with indented points.\n\nDesk feels quicker. Your message appears the moment you press Enter, and replies fade in faster.\n\nOne look across every panel. Line tabs, buttons that lift and press, switches that brighten on hover, a Config sidebar, centred pages, and a plain backup-answer note that names the model.\n\nConnectors load together instead of in two stages.\n\nThe VEKTOR Guide has five new chapters for the tools above.\n\n**Fixed**\n\nTrip card buttons now work when days only have labels such as “Thursday 17 Nov”.\n\nPage reading no longer fails when the browser engine is missing.\n\nThe image generator error that mentioned “mime” is gone.\n\nOpenverse errors say when the source’s own server is down, and the next press fails fast.\n\nRun your own race, keep your data sovereign.\n\nVersion v2.0.8 is live in the downloads page: https:[www.vekmem.com](http://www.vekmem.com)\n\nFull list of changes: [https://vektormemory.com/docs/changelog](https://vektormemory.com/docs/changelog)\n\nNovel Technology · Privacy Enhanced · Data Sovereignty · Vektor Memory\n\n600+ models BYOK · 60+ MCP tools · 16 Model Providers", "url": "https://wpnews.pro/news/vektor-2-0-8-studio-timeline-faraday-security-longmemeval-89-6-improvements", "canonical_source": "https://dev.to/vektor_memory_43f51a32376/vektor-208-studio-timeline-faraday-security-longmemeval-896-improvements-35pn", "published_at": "2026-11-18 23:24:08+00:00", "updated_at": "2026-11-18 23:42:10.492056+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "ai-tools", "large-language-models", "agent-protocols"], "entities": ["VEKTOR", "Faraday", "LongMemEval", "MCP"], "also_reported_by": [], "takeaway": "Faraday now blocks MCP rug pulls by re-checking tool descriptions after approval, caps helper agents to the tools of the agent that spawned them, and re-screens memories saved under suspicious content, while broader history search lifted LongMemEval from 84% to 89.6%.", "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/vektor-2-0-8-studio-timeline-faraday-security-longmemeval-89-6-improvements", "markdown": "https://wpnews.pro/news/vektor-2-0-8-studio-timeline-faraday-security-longmemeval-89-6-improvements.md", "text": "https://wpnews.pro/news/vektor-2-0-8-studio-timeline-faraday-security-longmemeval-89-6-improvements.txt", "jsonld": "https://wpnews.pro/news/vektor-2-0-8-studio-timeline-faraday-security-longmemeval-89-6-improvements.jsonld"}}