AGENTS.md turns transcript evidence into bounded agent memory

Kun Chen proposes treating a project AGENTS.md as adaptive memory: use agent transcripts as evidence, batch recurring failures, make small budgeted edits, and move narrow instructions into triggered skills. Why it matters: Agent instruction files usually grow through one-off reactions. A transcript-backed review loop makes changes easier to audit, challenge, and reverse.

Try this: Once a week, review a batch of agent sessions. Change no more than five rules, attach repeated evidence to each, and remove or extract an older rule to stay within a token budget.