A 435-paper survey says LLM agents still underbuild rollback, audit, and recovery

rohanpaul_ai · x · 2026-07-27

A survey argues that always-on agents should be treated as persistent-state systems, not just chatbots with memory.
The state they accumulate includes not only facts and preferences, but also permissions, credentials, task ledgers, triggers, provenance, and external commitments.
Across a scoped corpus of 435 works, retrieval appears in 269, writing in 200, rollback mechanisms in only 27, and authority in 72.
The paper says the field has focused much more on the forward path of writing/retrieving/acting than on the return path of forgetting, auditing, and recovery.
It proposes an Always-On Evaluation Protocol (AOEP-v0) to make governance requirements concrete by scoring state mutation and recovery obligations, and connects the agenda to databases, distributed systems, formal methods, capability security, and machine unlearning.

More from coding & agent