Building Agent M

In the last post I shelved a trading-engine rewrite and said the thing that actually needed rebuilding was how I work. This is that rebuild. I called it Agent M.

The first problem was the one I’d been circling since Sherwood: an agent that starts every session from zero. I wanted to give it what a fresh chat can’t have — a memory that carries across sessions and compounds the longer we work together.

The Karpathy wiki sketch from the last post was the shape of it: not a bigger context window, but a memory the model keeps and sharpens as it works.

At the center is that memory: a vault Agent M reads at the start of a session and writes back to as it learns, so the next one begins where the last ended. It didn’t start there — the first versions just kept a project’s notes, and the durable memory came a few weeks in, once it was clear how much those notes were already saving me.

Around the memory I wired the way I actually work — planning, reviewing, releasing — into steps the agent runs the same way every time, so I spend my time approving work instead of describing it. And I made it portable: I move between Claude Code and Antigravity depending on the task, and sometimes I ask Agent M for help while I’m on my phone, so Agent M runs the same on all of them. The memory and the workflow come with them, automatically.

It came together fast, the way these things do now. A rough version the same week I archived Sherwood, a real one by mid-May, and a steady run of releases since — it’s on its third major version as I write this. What started as a way to stop repeating myself became the tool I build everything else with.

The agent I work with now remembers. That’s the difference — and it’s the tool behind almost everything I’ve made since, including this blog.