Memory Management

This implements a three-tier memory system for AI agents: hot memory in MEMORY.md with P0/P1/P2 tags and TTLs, cold searchable storage in JSONL lesson files, and raw daily logs. The hook is the 70-80% token reduction claim, which makes sense since you're only loading ~200 lines of current context instead of everything. Comes with a Python janitor script that auto-archives expired entries based on priority and date tags. The discipline of deciding what's P0 versus P2 is actually the hard part here, not the tooling. If you're running long-lived agents that accumulate context bloat, this gives you a structured way to prune without losing important stuff to semantic search.