Layered Mutability: Five-Layer Governance Framework for Persistent Self-Modifying LLM Agents

arXiv·medium signal

A new framework (arXiv 2604.14717) introduces 'layered mutability' for reasoning about governance in persistent LLM agents across five layers: pretraining, post-training alignment, self-narrative, memory, and weight-level adaptation. The central claim is that governance difficulty increases with layer depth and that current oversight mechanisms don't account for mutable internal states influencing future behavior. For anyone building agents with persistent memory and self-reflection (MindPattern-style systems), this provides the first formal vocabulary for reasoning about which layers of agent state need governance controls.