I do not like them in my code. I do not like them, token load. I do not like re-reading files. I do not like re-learning styles. So I built HAM (Hierarchical Agent Memory).

Every Claude Code session, your agent burns 5,000–15,000 tokens just getting oriented. Re-reading your project. Re-discovering conventions. Re-proposing decisions you already rejected. One dev had 76K in a single CLAUDE.md eating 47% of his context window. Every request.

HAM replaces one bloated file with scoped memory.

Before: 5,000–15,000 tokens/request. After: 300–800. The system is self-maintaining. Agent creates memory files as you build.

Zero dependencies. Open source. For token-poor builders and small teams. Would you eat them in a box? Would you eat them with a fox?

No. You'd go HAM.