Learning

It is off until enabled in /admin/api/agent-learning/settings. The worker checks every minute and reviews only when no foreground task is active, at least five minutes after the last observation and an hour after the last review, at most three times per UTC day. A batch holds at most 20 experiences from a single audience and 48 KiB of evidence, so evidence never crosses between audiences.

The reviewer answers with a closed schema: keep, create, revise, replace, retire or soul, citing 1 to 20 evidence IDs, and an independent assessment turns an unsupported change into keep. At most 20 learned skills are active, each capped at 8 KiB and 300 lines. Soul revisions are appended to .blueclaw/state/persona/soul-history.json. Operators list, retire, restore and protect learned skills and read the soul history under /admin/api/agent-learning/*.