The knowledge layer for LLMs
All your AIs,
one brain you can read.
Vendor memory is scraped from your chats, locked to one product, and invisible to you. vtriv is the opposite. You decide what's worth keeping, it's stored as plain markdown, and one MCP connection puts it in front of Claude, Claude Code, or any other client.
Product engineer. Ships small, iterates in public.
## Stack
TypeScript and Go. Postgres, Redis, a bit of Rust. Boring tools on purpose.
## Working style
Concise. Bias to action. Pure functions, standard libraries,
no speculative abstraction. Tests first when it warrants it.
## Standing preferences
Commit completed work without being asked. No emoji in output.
inbox/ 2 unfiled captures
notes/
reading.md
ideas.md
errands.md
writing/
voice.md
pitch-notes.md
skills/
changelog/SKILL.md
release-notes/SKILL.md
projects/marlow/
README.md · notes/ · tasks/ · files/
Every line above is a file you can open, edit, or delete. Nothing is inferred from a chat you forgot you had.
Four things your assistant starts with.
Connect once and the whole layer is in scope. Your assistant reads it, and writes back to it, over the same connection.
Profile
Who you are, your stack, how you like to work. Always in scope, in every session, on every client. You write it once.
Skills
Agent Skills your assistant loads when the work calls for them. Keep them global, or wall them to a single space.
Knowledge
A markdown tree your assistant keeps in order. Plain files you can open, edit, and search. Nothing is hidden from you.
Projects and tasks
Bodies of work with tasks, notes, comments, and a timeline. The current state of what you're building, ready the moment your assistant needs it.
Memory is half of it. The other half is work.
A project carries tasks. Assign one to your agent and it picks the work up and runs with it. So vtriv isn't only what your AI remembers — it's where you hand it what to do.
labels marketing, design
blocked_by task:019f7ab2 (done)
Marketing site: new visual identity
— thread —————————————————————————
you Keep the copy. Restyle it as kept records.
claude Reading skill:brand-voice and skill:web-copy
first, then the site's styles.
claude Landed. Serif display, paper ground, one
accent. Build green, docs updated.
A real queue
A task carries an owner, a priority, a due date, a checklist. Chain them so the work happens in the right order, and the ones that repeat come back on their own.
Comments as conversation
Every task has a thread. You comment, the agent comments back, and the whole exchange stays attached to the work instead of scrolling away in a chat window.
Told when it matters
Watch a project and hear about it the moment a task moves or a comment lands. On your phone too, so you can step away from the desk and still know where things stand.
How it works.
Write down what's worth keeping
Fill in your profile, add a skill or two, put knowledge in the tree, start a project. Do it in the web app or just tell your assistant to do it. Keep separate spaces when work and personal shouldn't mix.
Connect once over MCP
Point any MCP client at https://mcp.vtriv.com/mcp and sign in through the browser. One global connection covers Claude, Claude Code, Cursor, and everything else — no per-project setup, no API key in a config file.
It reads, works, and writes back
Your assistant pulls the context a task needs, does the work, and files what it learned back into the layer. Next session starts from there instead of from zero.
Built for how a model actually reads.
Ask vtriv something and it doesn't hand back a pile of matches to wade through. One pack call gathers the most relevant passage from each hit — across your profile, skills, knowledge, and projects — and returns them as one bundle, already trimmed to fit what the model can read at once.
Every passage arrives labelled with where it came from, so your assistant can pull the one section it needs instead of a whole document. It spends its attention on your work, not on hunting for the six lines that mattered.
Deploys ship on merge. To roll back, redeploy the
last tagged release — never hotfix straight to prod.
file:notes/incidents.md#slow-reads
Check the cache first: a cold node drags every
request and shows up as slow page loads.
project:marlow#setup
docs/setup.md — build steps, env config.
A ChatGPT memory alternative you can read
ChatGPT's memory lives inside one product and you can't edit it as a file. The portable alternative: plain markdown every AI client reads and writes over MCP.
How to give Claude persistent memory across sessions
Claude forgets everything when the chat ends. Connect it to a context layer over MCP and every session starts knowing your stack, your style, and your work.
The cost of starting over
Every new chat begins from zero, so you retype the same background. A calmer way to work with AI assistants that don't remember you between sessions.
Your AI should already know you.
Free to start. Build the layer once and every assistant you use starts the session already knowing you.
Get Started