I built a self-improving AI agent. Started October 2025.

Six months in, the layer that actually made it learn was not memory. It was a small corrections loop with four sinks behind it. Capture → classify → graduate, drained every night. Plus a card in Basecamp for every correction, so I can see what the agent is teaching itself. 30 days in: 22 corrections, trending to 18.

New build log on the whole architecture.