Layered Recall

Progressive memory system with two orthogonal dimensions of lazy loading:

1. Scope layers - What is relevant (identity, project, domain, deep)
2. Depth layers - How much detail to fetch (IDs, summary, full)

Combined savings: 10-50x tokens vs eager loading.

Depth Pattern (Fetch-on-Confirmation)

Instead of loading full memory entries upfront, agents fetch in 3 depths:

Depth 1: IDs only        (~10 tokens per match)
  Agent decides which are worth investigating

Depth 2: Summary         (~50 tokens per match)
  Room, type, preview (first 80 chars)
  Agent confirms relevance