AI systems that compress old conversations to save memory can ace today's question and still get tomorrow's wrong, according to a new arXiv paper.
The researchers built what they call a paired-history audit: two different conversation histories that both produce the same correct answer right now, then receive the same new piece of information, and then require different correct answers afterward. In a pilot test, they ran 24 such history pairs across six synthetic memory-compression mechanisms, 12 memory conditions, and two AI model backends, DeepSeek and GLM. A deterministic "frontier" selector got every case right on DeepSeek but only 82 of 96 on GLM; a rival "structured writer" approach did worse still. Closer inspection found dozens of memory records that looked well-formed but were factually wrong, deleting an old entry broke every later reference in one setup (16 of 16), and simply renaming an identifier collapsed the system's accuracy from a perfect record to under a third.
That collapse is the real story here. Plenty of AI products now trim or summarize chat history to cut costs, on the assumption that a system answering correctly today is a system working correctly. This paper shows that assumption is shaky: the memory can look fine on every test you run against it now and still be quietly corrupted the moment new information needs updating.
The authors are upfront that this is a narrow, synthetic test, not proof that one memory scheme beats another, and none of it has been checked against real-world tasks. Read it as an early warning, not a verdict.