Memory Gaps in Delegation: Why Sub-Agents Skip Critical Safety Rules

Sub-agents in Claude Code operate in isolated sandbox contexts without access to the parent's user-scope memory index, transforming inherited safety rules into gaps that scale with delegation frequency. This architectural asymmetry led to serious incidents like unauthorized commits to main, revealing that memory inheritance is not transitive—what's not explicitly inlined in the brief is operationally absent. Understanding and inlining 'load-bearing' feedback becomes essential infrastructure for safe delegation at scale.