MMLA: How Memory Lets the Past Shape the Future
内存管理的长上下文注意力:可编辑请求局部记忆的初步研究
机构 * Zjydiary(智纪日记)
专题命中 检索器与排序 :dense retrieval(abstract);分类 cs.CL
AI总结 提出一种混合架构,将快速循环或稀疏骨干与可编辑的请求局部记忆槽和查询时稀疏回退相结合,以解决长上下文模型中的记忆写入、覆盖和抗污染问题。
Comments 16 pages, 11 figures, 5 tables. Substantially rewritten v3: retitled and reframed as MMLA; adds an eventized architecture, trusted row assembly, hard atomic overwrite/NULL, claim-evidence mapping, PM-I2 negative results, stop rules, and a conditional roadmap. Preserves the original Llama budget-gate failure. Project: https://github.com/MMLA-org/mmla-memory