Fresh Memory, Stale Plans: Dependency-Scoped Validation for Distributed LLM-Agent Memory
Evan Chen, Shiqiang Wang and Christopher Brinton (Purdue and Exeter) name stale-plan execution, where a distributed agent team reads perfectly fresh shared state and still acts on a plan derived from a requirement that has since been superseded.
Ask this paper
Freshness is not validity: a planner derives an action from requirement r3, another agent commits r4, and the executor receives r4 without replacing the r3-derived plan. State freshness never established that the authorizing plan still holds.
PlanFence cites its sources: plans record the exact public records they used, and the executor validates only the records that can affect the pending external action, then replans once or blocks if validation is incomplete.
Total in the controlled setting: across 30 live workflows with a post-plan revision, a freshness-only executor acts on the obsolete plan in every single task while PlanFence completes all of them with no invalid action.
Two conditional boundaries reported honestly: proactive synchronization has lower stall at low churn, while PlanFence wins as churn grows and as the shared keyspace grows.
Scoped claim: the authors state these are controlled safety and systems-cost results, not general task-accuracy gains.
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 $r_3$, another agent may commit $r_4$, and an executor may receive $r_4$ without replacing the plan derived from $r_3$. We call this \emph{stale-plan execution}: state freshness does not establish that the plan authorizing an action remains valid. We introduce PlanFence, a dependency-scoped action-validation protocol. Plans cite the exact public records they used, and an executor validates only the records that can affect the pending external action, replanning once or blocking when validation is incomplete. In 30 controlled live workflows with a post-plan revision, a freshness-only executor acts on the obsolete plan in every task, whereas PlanFence completes all tasks without an invalid action. Controlled replay reveals two conditional boundaries: proactive synchronization yields lower coordination stall at low churn, while PlanFence avoids repeated update-path coordination as churn grows and avoids validating unrelated state as the shared keyspace grows. These are controlled safety and systems-cost results, not general task-accuracy gains.