A new study on AI agent memory finds that the exact wording of a stored instruction, not just its content, can steer which piece of memory an agent retrieves and acts on.
Researchers ran twelve registered experiments across 14,760 attempts, testing agents that inherit a handful of short memory entries and must choose one archived record before acting. They compared three ways of directing that choice: a simple pointer to the record, a written criterion describing which record to pick, or both together. A criterion beat a bare pointer by a wide margin on six directly-hosted models, though that advantage did not hold up on a separate nine-model panel accessed through a routing service. On Claude's Opus and Fable model lines, appending an identifier after a criterion could cancel the criterion's effect entirely, sometimes flipping an agent's behavior from always following the rule to never following it. Adding an explicit confirmation line and a small compute budget restored the intended behavior in every case tested.
This matters because it is not really a story about model capability. It is a story about how fragile instruction-following can be at the exact moment an autonomous agent decides what to act on. The same underlying instruction produced opposite results depending on formatting choices that a human reviewer might consider cosmetic, which is an awkward finding for anyone building agents that manage their own memory over long-running tasks.
The researchers are careful to call these descriptive effects on fixed test setups, not evidence of any underlying mechanism, and that caution seems warranted. But for teams shipping memory-augmented agents today, the practical takeaway is blunt: test your prompt formatting like it is application logic, because right now it behaves like one.