What is HAM?
HAM (Hierarchical Agent Memory) is a Claude Code skill that sets up scoped CLAUDE.md files per directory, reducing agent context token spend by 80-90%.
Instead of one massive root CLAUDE.md loaded on every request, HAM creates a tree of lightweight context files — each containing only what the agent needs for the directory it's working in.
Key features:
- Auto-detects platform (Web, iOS, Android, Flutter, React Native, Python, Rust, Go) and project maturity
- Three setup paths: Greenfield, Early Stage, and Brownfield
- Self-maintaining — operating instructions tell the agent to update memory as it works
- Inference quarantine — uncertain discoveries go to inbox.md, never directly to canonical memory
- Proactive audit suggestions after 10 sessions or 14 days
- Token savings tracking with transparent calculations
Repo: https://github.com/kromahlusenii-ops/ham