What problem does it solve? Your agent forgets everything between sessions, so recurring mistakes and preferences never turn into lasting improvements. This Skill runs an offline sleep cycle that reviews past Claude Code sessions, replays recurring tasks, and consolidates validated learnings into memory and skill files. ## Core Features & Use Cases - Six-stage sleep cycle: harvest past session transcripts, mine recurring tasks, replay them offline, consolidate bounded edits, stage proposals, and adopt only after a held-out validation gate passes. - Gated memory and skill evolution: proposes bounded add/delete/replace edits to CLAUDE.md and SKILL.md, accepting changes only when they strictly improve held-out scores, with backups before adoption. - Scheduling and safe preview: install a nightly cron entry, run a zero-cost mock backend dry-run, or spend real API budget with claude/codex backends. - Use Case: Schedule a nightly run so your agent reviews the day's sessions at 3 AM, stages a proposed CLAUDE.md update encoding your preferences, and shows you the held-out score improvement before you adopt it. ## Quick Start Ask your agent to run the skillopt-sleep dry-run on the current project to preview what a nightly self-improvement cycle would propose.