Fresh Memory, Stale Plans: PLANFENCE Blocks All Invalid Agent Actions Across 30 Workflows
rohanpaul_ai · x · 2026-09-09
A new paper shows that refreshing an agent's memory doesn't refresh its plan: in all 30 live workflows with a freshness-only setup, the agent saw updated records but still issued obsolete actions.
PLANFENCE's approach:
- Ties each plan to the exact records it used and checks only the records relevant to the next tool call
- If something changed, the agent replans once; if the plan still can't be verified, the action is blocked
- Result: every invalid primary action across the 30 workflows was avoided
On performance, proactive sync was faster at low update rates, but PLANFENCE had lower stall as updates increased, while avoiding checks on unrelated shared state.
Key takeaway: fresh memory is not proof that a plan is still valid — re-validate plans before execution. Paper: arxiv.org/abs/2609.03340.
More from coding & agent
- Nous Research pitches Hermes Agent as the only way to truly own your AI stack — Teknium · 2026-09-09
- One Person + Codex + Apify + Apollo: The Exact Stack for AI-Powered Outbound That Isn't Spam — aryanXmahajan · 2026-09-09
- GitHub Details 4 Changes That Cut Copilot's AI Cost Without Hurting Task Quality — marlene_zw · 2026-09-09
- Mac MCP 2.0 open-sources 81 tools giving AI agents local macOS control and background browser automation — bulutarkan · 2026-09-09
- Teknium: Ran Hermes 19 Hours with 1,390 Subagents and Zero Failures — Teknium · 2026-09-09
- Lessons from building a local macOS MCP server for ChatGPT-assisted coding and admin work — bulutarkan · 2026-09-09