Original title: Plan Pointers and Record-Directive Form in Budgeted Verification of Inherited Agent Memory
Kazuki Nakayashiki · September 2026
Artificial intelligence agents sometimes inherit stored memories and have a budget allowing them to check only one source record before making a decision. Researchers wanted to know how different written instructions steer which record the agent inspects. They tested instructions written as a direct pointer, a descriptive criterion that identifies the record, or both together. Across thousands of tests, they found that writing out a descriptive criterion worked much better than giving a bare identification code on several models. Surprisingly, adding the identification code after the criterion cancelled out the criterion on specific models, dropping success to zero. However, adding an approval line or increasing the budget to check two records restored the desired choice. Finally, these instructions also shifted the agents' ultimate choices, though different models reacted in opposite directions. Tiny phrasing changes created distinct outcomes.
Imagine you're choosing a book from a library cart based on a sticky note. If the note says 'the red book,' you'll grab it quickly. But if the note says 'a book about dogs,' you might pick a different one. Now suppose someone adds 'the red book' to the note that says 'a book about dogs'—suddenly, you ignore the dog part and just grab the red book. In this paper, the studies show how such small changes in the instruction affect which record the AI picks, mirroring how the note's wording changes your choice.
This matters because developers building AI agents need to know that tiny tweaks in instructions can make or break performance. If a system is sensitive to small changes, it might behave unpredictably in real-world use, affecting anyone relying on automated decisions.
The results are descriptive effects on fixed panels with no mechanism claim, meaning it doesn't explain why these changes happen. The studies are limited to specific AI models and instruments, so they may not generalize to all systems.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Plan Pointers and Record-Directive Form in Budgeted Verification of Inherited Agent Memory — Kazuki Nakayashiki
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.