Five AI agent harnesses took the same memory test. The results weren’t even close. The New Stack