When LLMs Fall into Forgetting, Where Do the Memories Go? A Cost-Compressed, Ultra-Lightweight Biomimetic Memory Retrieval Engine and Fast Edge Search Architecture (当大语言模型坠入遗忘之渊,记忆流向何方?)

【实测峰值验证数据 (Peak Verified Performance)】Funes 核心引擎常驻内存仅 18.6 MB,GPU 显存占用绝对为 0%,记忆写入时消耗 0 Token,检索召回时消耗 0 Token,端侧单次检索典型延迟低至 6.6 ms。在 ICLR 2025 权威长程记忆基准 LongMemEval-S 全量 500 题严格物理双盲评测中,全流程仅耗时 71.36 秒(峰值吞吐达 9.5 题/秒),实现 97.80% (489/500) 的任意目标会话命中率(Any-Session Hit Rate)、92.00% (460/500) 的完整证据全召回率与 81.60% (408/500) 的黄金答案直接子串命中率,端到端问答准确率达 93.20%*。 Long-term memory for Large Language Models (LLMs) remains severely constrained by dense vector embeddings and graph indexing. Biological human memory operates with astonishing metabolic frugality. We present Funes (HLM), an ultra-lightweight, zero-embedding biomimetic dual-stream memory engine operating natively on the CPU with zero external dependencies. Funes maintains a resident footprint of 18.6 MB RAM, 0 MB GPU VRAM, 0 tokens ($0.00 compute cost), and point lookups down to 6.6 ms. On the full 500-question ICLR 2025 LongMemEval-S benchmark, the entire suite executes in just 71.36 seconds, achieving a 97.80% Any-session retrieval hit rate (489/500), a 92.00% Full-session recall rate (460/500), and an 81.60% direct ground-truth substring match rate (408/500).