Fresh Memory, Stale Plans: PlanFence — Dependency-Scoped Validation for Distributed LLM-Agent Memory

Forum topic · 小凯 · 2026-09-06

Summary

This arXiv paper (2509.00004) by Evan Chen, Shiqiang Wang, and Christopher G. Brinton addresses a subtle failure mode in distributed LLM-agent teams called stale-plan execution: agents can read the latest shared facts yet still act on an obsolete plan. In the failure scenario, a planner derives an action from requirement r3, another agent commits r4, and the executor receives r4 without replacing the plan derived from r3 — so state freshness alone does not guarantee the authorizing plan remains valid. The authors introduce PlanFence, a dependency-scoped action-validation protocol in which plans cite the exact public records they used, and executors validate only records that can affect the pending external action, replanning once or blocking when validation is incomplete. In 30 controlled live workflows with post-plan revisions, a freshness-only executor acted on obsolete plans in every task, while PlanFence completed all tasks with no invalid actions. Controlled replays also show that proactive synchronization yields lower coordination stall at low churn rates, while PlanFence avoids repeated re-plan coordination as churn grows and avoids validating unrelated state as the shared key space grows. The authors emphasize these are controlled safety and system-cost results, not general task-accuracy improvements.

Overview

Field: AI/ML Authors: Evan Chen, Shiqiang Wang, Christopher G. Brinton Published: 2026-09-06 arXiv: 2509.00004

The Problem: Stale-Plan Execution

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 r₃, another agent may commit r₄, and an executor may receive r₄ without replacing the plan derived from r₃.

The authors call this stale-plan execution: state freshness does not establish that the plan authorizing an action remains valid.

PlanFence

PlanFence is a dependency-scoped action-validation protocol:

Plans cite the exact public records they used.
An executor validates only the records that can affect the pending external action.
When validation is incomplete, the executor replans once or blocks.

Key Results

In 30 controlled live workflows with a post-plan revision, a freshness-only executor acted on the obsolete plan in every task, whereas PlanFence completed all tasks with no invalid actions.
Controlled replays reveal two conditional boundaries:
Proactive synchronization produces lower coordination stall at low churn rates.
PlanFence avoids repeated re-plan coordination as churn rate grows, and avoids validating unrelated state as the shared key space grows.

Scope Caveat

These are controlled safety and system-cost results, not general task-accuracy improvements.

*Auto-collected on 2026-09-06.*