Why I started Agent M

The hole in every agent

Every coding agent I tried had the same hole in it. It was brilliant for an hour and a stranger the next morning — no memory of the decision we had locked, the convention we had agreed, the bug we had already fixed twice.

Building the scaffolding

So I stopped waiting for a smarter model and started building the part everyone skips: not the intelligence, the scaffolding around it.

“Useful enough to depend on” and “safe” turn out to be the same problem.

The first thing it needed was memory that survives the session — and a way to write to it I could trust. So the harness gates on it:

$ agentm /plan "build the blog"
recall vault state ........ ok (3 always-load, 2 project)
no plan in flight ......... proceed
wrote PLAN.md ............. 7 tasks, smallest-first

What memory actually needs

A short clip of the harness recalling state at the start of a session (placeholder embed — it loads only when you click):

Seed post — the full essay develops from the Agent M Evolution HLD.