HAM, Hierarchical Agent Memory

A memory system for AI coding agents. Scope, don't dump.

A memory system for AI coding agents, born from one annoyance: every session re-reads everything, and a bloated context file can eat half the window before any code gets written.

HAM had one idea — scope, don’t dump. Small context files at each directory level instead of one giant one at the root: global conventions in root, API patterns in the API folder, component rules with the components. In my testing that cut context per prompt from 4,000–12,000 tokens to 800–2,400 — roughly an 80% reduction, with a savings report and even a carbon estimate built in.

MIT licensed and still up. Take it, fork it, gut it for parts.