# PLANFENCE：验证智能体记忆与计划一致性

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-09-09 17:46
- AIHOT 分数：46
- AIHOT 链接：https://aihot.news/items/cmttxtrs90rimrofpbm9batrg
- 原文链接：https://x.com/rohanpaul_ai/status/2097622540591960130

## AI 摘要

论文显示，刷新智能体记忆并不会刷新其计划，在30个实时工作流中，仅刷新记忆的设置下智能体均看到新记录却仍执行过时操作。PLANFENCE通过将计划与所用记录绑定并仅检查影响下一步调用的记录来解决此问题，若变更则重新规划一次，无法验证则阻止操作，成功避免了所有无效主操作。

## 正文

The paper shows that refreshing an agent's memory does not refresh its plan, so the plan should be checked again before execution.

One agent can update a requirement, and the executor may see that update while still holding a plan built from the older version.

That happened in all 30 live workflows with a freshness-only setup: the agent saw the new record but still issued the obsolete action.

PLANFENCE fixes this by tying each plan to the exact records it used and checking only the records that matter to the next tool call. If something changed, the agent replans once; if it still cannot verify the plan, it blocks the action.

That avoided every invalid primary action in the 30 workflows.

At low update rates, proactive sync was faster. As updates increased, PLANFENCE had lower stall, and it also avoided checking unrelated shared state.

So, do not treat fresh memory as proof that an agent's plan is still valid.

– arxiv. org/abs/2609.03340

Title: "Fresh Memory, Stale Plans: Dependency-Scoped Validation for Distributed LLM-Agent Memory"