continuous-learning-v2

To turn Claude Code sessions into reusable knowledge, this Skill observes raw session data and converts it into atomic, confidence-weighted instincts. These small behaviors are stored with metadata for future reuse across projects.

Atomic instincts are small, testable behaviors extracted from session observations. They feature confidence scores and frontmatter-driven metadata, providing a structured, repeatable method to scale knowledge growth across projects.

You can evolve project-scoped instincts into global Claude Code commands through optional cross-project promotion. The system stores data under ~/.claude/homunculus and supports promoting local behaviors into skills, commands, or agents.

No external dependencies are required to extract knowledge and generate instincts. The Skill operates independently using internal scripts, storing all persistent data locally under the ~/.claude/homunculus directory with project and global scoping.

You automate continuous learning by integrating the Skill's hooks and CLI into your existing Claude Code workflows. This setup continuously observes sessions, generates new instincts, and evolves existing behaviors without manual intervention.

The best way to isolate project knowledge is using the default project-scoped storage. The Skill enforces metadata like project_id and scope, keeping instincts deterministic and isolated within the ~/.claude/homunculus directory structure.