A structured .json or .md artifact generated by complex skills (e.g., agent-browser) at the end of their execution. It explicitly logs the agent's goal, intermediate tool calls, and final success/failure state, serving as the raw input (memory) for background reflection. Stored temporarily in .tmp/traces/.
An asynchronous batch-processing script (or skill) that runs during downtime (e.g., as part of the daily-workflow). It consumes accumulated Trace files, evaluates them using an LLM to identify recurring errors or anti-patterns, and extracts permanent rules to prevent future failures.
The process of permanently storing the rules extracted by the Dreamer. To prevent global context bloat, the rule is appended locally to the specific skill's SKILL.md (e.g., under a ## Known Gotchas & Lessons section). Concurrently, the skill's README.md changelog must be updated to reflect the newly injected rule.