When Memory Becomes Authority: Benchmarking Authority Collapse at the Memory Consolidation Boundary
2026-08-03 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors study how AI agents with persistent memory store facts and rules from past interactions to improve future tasks. They find a problem called authority collapse, where the system forgets the rules about how stored information is allowed to be used, leading to incorrect or unauthorized actions. They create a test called AuthMem-Bench to measure this issue and show that most current systems suffer from it. By adding metadata that keeps track of usage permissions, the authors greatly reduce unauthorized actions without hurting task performance. This means AI memory needs to remember not just facts, but also the authority rules for using them.
Persistent memoryLLM agentsAuthority collapseInformation consolidationAuthorization boundaryAgent-memory systemsUnauthorized actionsMetadataAuthMem-BenchTask adaptation
Authors
Qiuyang Zhan, Rui Zhang, Sheng Guo, Lepeng Zhao, Zhuotao Liu
Abstract
Persistent memory allows (self-evolving) LLM agents to adapt across tasks by consolidating heterogeneous interaction histories into reusable facts, preferences, observations, and rules. Yet consolidation also imposes an implicit authorization boundary: it determines whether stored information may later be consumed as a user fact, an attested observation, or a standing instruction. We identify authority collapse, in which consolidation preserves a claim while erasing the source constraints governing its authorized use, causing the stored memory to imply greater authority than its source permits. We introduce AuthMem-Bench, a controlled paired benchmark that holds the focal claim and downstream task fixed while varying only source authority. It evaluates write-time collapse, downstream authorization errors, and automatic authority preservation. Across seven consolidators based on widely used agent-memory systems and seven LLM backbones, we observe authority collapse in 48 of 49 evaluated configurations. In a controlled action-grounded evaluation, collapsed memories without authority metadata yield a mean unauthorized-action rate of 50.3%. In an end-to-end evaluation, automatically predicted and persisted authority labels reduce the observed unauthorized-action rate from 16.9% to 0.0%, while benign task success remains essentially unchanged. These findings show that memory-driven adaptation must preserve not only what was learned, but also the authority under which it may be reused.