Study: Mid-Tier AI Helpers Are Most Likely to Trust Outdated Information

AI systems are increasingly built like a small office. One AI acts as the manager, then splits the work and sends out helper AIs — often called "sub-agents," meaning AIs that can go off and do tasks on their own. By default, most of these setups hand the helper the manager's entire memory: the full conversation, every conclusion reached, every dead end. Researchers at a multi-agent systems lab asked a simple question — does that inherited memory help or hurt? They designed a test where every task could be solved from the original evidence alone. So any drop in accuracy had to come from trusting old notes.

The surprise was who failed. You'd expect the weakest AI to be the most confused, and the strongest to stay sharp. Instead, the middle-strength model was the worst off. A mid-tier model scored about 19 points lower than when it started fresh, and did worse than a weaker model and a stronger one. The researchers call this the "danger band." The comparison: a junior employee who understands instructions well enough to follow them, but not well enough to stop and ask, "Wait, is this memo still true?" The weakest AI barely followed the old notes at all. The strongest spotted the contradiction and ignored them.

This matters because AI assistants now handle email, refunds, bookings, research and customer service. Those helpers often inherit context like "this customer was already refunded" or "the flight was moved to Tuesday." If that conclusion is out of date, the helper can repeat the error confidently — and the person on the other end may never know why. The paper's practical recommendation is straightforward: send a short, hand-picked summary rather than the full history. That approach beat the "send everything" default on all three datasets, with the biggest gain at the mid-tier model. Simple rules like "only share detailed notes with bigger models" failed when tested.

One honest caveat: this is a preprint, not yet peer-reviewed, and it ran on frozen lab benchmarks rather than real customer systems, so the exact size of the effect in the wild is unknown. Still, as more companies ship AI helpers that act on your behalf, it's a useful warning: more shared memory isn't always better memory.