hierarchical-agent-memory

You can reduce token waste by implementing directory-scoped CLAUDE.md memory files. This approach provides sub-context for specific directories so agents load only the relevant context instead of rereading entire project structures repeatedly.

Directory-scoped memory is a context routing technique using root and per-directory CLAUDE.md files. You need it for multi-directory projects where agents repeatedly read the same files, causing unnecessary token consumption and slower response times.

Run HAM setup to auto-detect your project structure and generate CLAUDE.md memory files. This creates a root file for global context and subdirectory files for scoped context, then launches a dashboard to visualize token savings.

Yes, you can track token savings using a lightweight .memory dashboard. It visualizes token savings and context health while providing an audit trail to monitor directory-scoped memory decisions across your project.

Yes, directory-scoped context routing works well for onboarding repos with deep directory structures. It improves response time and reduces prompts by ensuring agents load targeted sub-context rather than scanning the entire project.

This approach requires maintaining both a root CLAUDE.md and per-directory files, which adds setup overhead. It is best suited for multi-directory projects where agents repeatedly read identical files and prompts.