New AI Trick Helps Research Bots Remember Their Own Search Steps

When you ask an AI a big, complicated question — like "What's the best electric car for a family under $40,000?" — it doesn't answer in one go. Instead, it runs a series of smaller searches, gathering pieces of information and then combining them. These are called "agentic search" systems. But there's a problem: most AI training treats each of those small searches as if they were brand new, ignoring what the AI already learned in earlier steps.

Researchers from Australia, China, and Thailand created ITER, a new training method that fixes this. Think of ITER as giving the AI a memory of its own homework. As the AI searches, it learns not just from the current question and results, but also from the original big question and all prior sub-questions it asked. This lets the AI make smarter decisions about what to look at next — and avoid re-reading pages it already visited.

The results were strong. Across six different AI systems, ITER improved accuracy by 7.5% on one research benchmark and 13.5% on another, compared to current best methods. It also worked well when paired with AI systems it had never seen before, which suggests it's a flexible upgrade rather than a one-trick fix.

What does this mean for you? If you use tools like ChatGPT's deep research or similar AI-powered search assistants, systems like ITER could lead to answers that are more thorough, better sourced, and less likely to miss key details. In other words, AI research is getting closer to acting like a thoughtful human researcher who remembers the whole conversation — not just the last thing it read.