Noodle

Whenever you head into a meeting you bring a whole bunch of things with you. You know the strategy. You know the roadmap. You know what conversation you had with your designer last week. You switch between people and contexts and you're able to hold all that in your head.

But today, your AI agent doesn't do this. It starts every conversation fresh. You have to brief it with the context every time you talk to it.

What if it already knew?

This morning I opened Claude Code and asked it to start my day. It already knew where my team's GitHub boards were. Knew the state of our pipeline and new closed deals. Scanned my to do list to know what the most important things on the agenda are. Looked at competitors and their movements. It has a shared context that it creates, but that I build on by talking to it. It flags the things I forgot. It connects the dots I would connect if I had the headspace.

Context in Projects rots over time

The idea started way back when projects were released in the web interface. I'd dump notes in. Feed it context.

But the context rots. Because strategy isn't static. Because maintaining docs is a chore. And no one does that. The stuff that matters got buried in a chat from 4 weeks ago, and the stuff I'd put in the UI was old. Out of date.

But what if using it could be maintaining it?

At the back end of last year I heard a lot of noise about using markdown files and creating a memory that way.

I started with a briefing doc. Just a claude.md. That's not revolutionary. You tell it what matters. A short but insightful document that helps it orient itself.

Then you add the maintenance layer. Something that does a few things at once. It looks for docs you've updated recently. Stuff you don't look at fades. Things you look at more often get mentioned more often.

Next up. Usefulness. The system watches to see what happens after you look at a doc. Did it lead somewhere? Did you cite it? Come back to it? The ones that are useful get surfaced more. The ones that aren't. Don't.

Then you layer on a shared vault written in markdown. Decisions, meeting notes, strategy, product context.

Without this your AI is brilliant. But they're forgetful. You have to constantly re explain the same stuff over and over. Yes, projects help with this, but the context gets old and stale and less useful over time.

With this your AI shows up prepped and ready. Not because you've been hand crafting and maintaining a knowledge base, but because every time it does something it leaves the system a little sharper about what matters and what doesn't.

The context maintains itself just by you working with it.

How it works

This is how my home version of Noodle works. My work version is a little more locked down, with no email route in.

Credit

The scoring engine is built on ideas from Ori Mnemos - the Q-value reward signals, ACT-R decay, and retrieval fusion. I simplified the implementation and wrapped it in the workflows I actually needed like the email pipeline, overnight batching, the vault hygiene hook, and the multi-model review pipeline.