研究人员 Vignesh Kothapalli 及其团队于2026年11月3日发布了名为 Memorilla 的内存系统,旨在通过固定数量的“记忆令牌”增强冻结状态大语言模型(LLM)在处理长上下文及知识密集型任务时的表现 1。该系统的工作原理是将文档编码为可重复使用的嵌入向量,并在输入嵌入空间中检索与特定查询相关的信息,从而优化模型对大量数据的处理效率 1。
在性能基准测试中,Memorilla 展现出显著优势,其仅使用16个记忆令牌便在多项任务上超越了 Mem0 和 MemoRAG 等基线模型 1。具体数据显示,在 FactKG 基准测试中,Memorilla 达到了84.33%的准确率,高于全上下文推理模式的80.43% 1。此外,该系统在代表源文档时所需的 Token 数量减少了108倍,极大降低了计算成本 1。在 TextWorld 环境的实验中,Memorilla 使平均成功率从18.6%提升至24.2%,证明了其在提升智能体系统持续学习能力方面的有效性 1。
Researchers have introduced Memorilla, a memory system designed to enhance the capabilities of frozen large language models (LLMs) in handling long-context and knowledge-intensive tasks 1. The architecture operates by encoding documents into reusable embedding vectors and retrieving query-specific information within the input embedding space using a fixed number of "memory tokens" 1. Developed by Vignesh Kothapalli and their research team, the paper was published on November 3, 2026 1.
In benchmark evaluations, Memorilla outperformed baseline models such as Mem0 and MemoRAG across multiple tasks while significantly reducing token consumption 1. Specifically, the system achieved an accuracy rate of 84.33% on the FactKG benchmark, surpassing the 80.43% recorded by full-context inference 1. This efficiency gain is attributed to representing source documents with 108 times fewer tokens compared to traditional methods 1. Additionally, experiments conducted in the TextWorld environment demonstrated that Memorilla improved the average success rate from 18.6% to 24.2%, highlighting its potential for enhancing continual learning in agent systems 1.
评论
还没有评论,欢迎留下第一条。