Fresh Memory, Stale Plans: Dependency-Scoped Validation for Distributed LLM-Agent Memory
Source
arxiv.org
Author
Evan Chen, Shiqiang Wang, Christopher G. Brinton
Date
Context
An AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → can receive an updated requirement and still execute a plan built from the old one. PlanFence addresses that mismatch by recording which versions of shared records produced a plan. Immediately before an external action, it checks the relevant records, requests one replan if they changed, and blocks when validation remains incomplete.
In 30 controlled workflows where researchers inserted a revision after planning, PlanFence completed every task without an invalid primary action. An executor that merely read the latest requirement acted on the obsolete plan in all 30. This isolates a specific coordination failure; deployment still depends on correctly recording every dependency that can affect the action.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
articleEnvironment Steering: Using Data Flow Control to Improve Agent Utility and SafetyCharlie Summers, Prajwal Raghunath, Aaditya Pai, Mayur Kulkarni, Zhuo Zhang, Oliver Kennedy, Eugene Wu
articleCallability Is Not Operability: Controlled Interface Interventions for LLM AgentsZihao Wang
articleThe Memory Trust Gap: Capability-Dependent Failures in Persistent-Memory AgentsJundong Hu, Shekar Ramachandran