Researchers propose Always-On Agents for persistent AI systems

Researchers Tianyu Ding, Aditya Nannapaneni, Bingfan Liu, and Ling Zhang from the computer science community have published a comprehensive survey titled *Always-On Agents: A Survey of Persistent Memory, State, and Governance in LLM Agents* on arXiv (arXiv:2606.30306). The paper introduces a new framework for AI systems that maintain persistent state across interactions, treating them as durable systems where future behavior depends on accumulated memories, task ledgers, permissions, and audit records.

The survey proposes six diagnostic axes for state management—authority, scope, mutability, provenance, recoverability, and actionability—and outlines a lifecycle for state handling from writing to forgetting. Notably, the authors highlight a research gap: existing work focuses more on accumulating/retrieving state than on governance, recovery, or relinquishment. To address this, they introduce the *Always-On Evaluation Protocol (AOEP-v0)*, a pilot evaluation framework that prioritizes state governance over answer quality alone. The work connects persistent agents to databases, distributed systems, and machine unlearning, offering a roadmap for next-generation AI systems.