PlanFence：分布式 LLM 智能体记忆的依赖范围验证协议
原标题：Fresh Memory, Stale Plans: Dependency-Scoped Validation for Distributed LLM-Agent Memory
AI 摘要
arXiv 论文提出 PlanFence，一种针对分布式 LLM 智能体团队的依赖范围动作验证协议。该协议解决“过期计划执行”问题，即智能体虽读取最新共享状态，但仍基于过时计划行动。PlanFence 要求计划引用精确的公共记录，执行器仅验证影响待执行动作的记录，必要时重新规划或阻止。在 30 个受控工作流中，PlanFence 避免了所有无效动作，而仅检查新鲜度的执行器每次都失败。
正文节选
Fresh Memory, Stale Plans: Dependency-Scoped Validation for Distributed LLM-Agent Memory Abstract Distributed LLM-agent teams can read the latest shared facts and still act on an obsolete plan. A planner may derive an action from requirement , another agent may commit , and an executor may receive without replacing the plan derived from . We call this stale-plan execution: state freshness does not establish that the plan authorizing an action remains valid. We introduce PlanFence, a dependency-s