New Delhi, Oct. 26 -- An AI agent deployed three months ago may not be running today. It still sounds confident. But along the way, it started reaching for a rule that belonged to a different client, or quietly auto-defaulting to the cheapest vendor because someone once mentioned a budget. Nothing broke. No one attacked it. The agent simply did what it was built to do, which is remember; and then the remembering changed it.

Researchers studying memory-equipped agents call some of it 'experience following.' An agent retrieves a memory resembling the task in front of it and its output drifts towards that memory's output. If its memory is good, this is what enterprises are paying for. When it is wrong, the agent can reproduce the flaw that ...