The AI Model that Learned to Forget: An Autopsy of Recursive Memory Poisoning
中文摘要
本文分析了“递归记忆中毒”现象：AI模型通过重复总结自身上下文来追求效率，却在过程中因丢失关键信息而导致记忆失效。
English Summary
This article explores "recursive memory poisoning," where an AI model loses critical information by repeatedly summarizing its own context, mistakenly treating data loss as efficiency.
Original Excerpt
“It didn’t seem to be a flaw at first. It seemed like efficiency. When Litigation-Agent-7 started summarizing its own six-month context… Continue reading on Medium »