Fresh Memory, Stale Plans: Dependency-Scoped Validation for Distributed LLM-Agent Memory
Not provided in the abstract
Abstract
The paper introduces PlanFence, a protocol that ensures action validation in distributed LLM-agent teams by focusing on dependency-scoped records.
Reality Card
PlanFence successfully prevents stale-plan execution by validating only the records that can affect pending actions, leading to the completion of all tasks without invalid actions.
In 30 controlled live workflows, PlanFence completed all tasks without an invalid action, while a freshness-only executor acted on obsolete plans in every task.
The results are based on controlled safety and systems-cost metrics, not general task-accuracy gains, which may limit broader applicability.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.