Upgrade — Self-Improving Agent Loop

Your agent makes a mistake once. Never twice.

Agents have no memory between sessions. You correct the same mistake on Monday, Wednesday, and Friday. Upgrade gives your agent institutional knowledge — every failure becomes a permanent lesson.

5 capture triggers:

1. Command fails (e.g., pip on macOS)
2. User corrects the agent
3. Tool or capability missing
4. External API fails
5. Better approach discovered

Each lesson is logged to LEARNINGS.md with date, trigger, fix, category, and severity. Before starting major tasks, the agent scans past lessons and avoids known mistakes entirely.

Real production examples included: macOS pip3, API rate limit backoff, config backups before edits, cron model routing (prevented $50/day API charges), memory file size management.

Use cases: development workflows, API integration, system administration.

v1.1 — includes capture trigger details, lesson format spec, troubleshooting, and quarterly maintenance guide.