What's worth keeping for your AI
Once your assistant can read a context layer, the real question is what to store in AI memory. Four kinds that earn their place, and what to leave out.
Say your assistant can finally read a shared context layer — you’ve stopped starting over every session. Now the real question: what goes in it?
Most of what happens in a chat isn’t worth keeping. The throwaway questions, the half-formed tries, the stuff you’ll never look up again. Save all of it and you’ve built a junk drawer — which is roughly what a model does when it curates memory for you, and one reason to prefer a memory you write yourself. Search turns noisy, the assistant grabs the wrong thing, and the layer becomes one more mess to manage.
Four kinds of context earn their place.
Who you are. Your stack, your preferences, the constraints that hold week to week. Write it once. It stays in scope, so you stop opening every chat with “I’m a Rust developer who avoids frameworks.”
Reusable skills. The how-tos you’d otherwise repeat: the way you like commit messages written, the steps to cut a release, the house style for an API. Package it once. Any assistant can load it when the task calls for it.
Reference knowledge. The durable material your work leans on. Specs, research, the reasoning behind a design. The stuff you’d dig out of a folder, except the assistant reads it the moment a task needs it and writes new findings back.
Active work. What you’re shipping right now: projects, tasks, the running log of what changed. This context goes stale fastest and matters most while it’s hot.
Notice what’s missing. Chat logs. Every message you’ve sent. A full export of your notes app. More isn’t better here. A lean layer the assistant can search beats a sprawling one it drowns in.
Here’s the test: if you’d never look it up again, don’t save it. If you’d groan at re-explaining it, do.
vtriv gives each of these a home: profile, skills, knowledge, and projects with tasks. Your assistant reads across all of them over one MCP connection, and writes back as it learns — connecting it takes a single command. Start here.