<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>上下文压缩 on Architecting Intelligence</title><link>https://blog.oklensuangoo.uk/zh/tags/%E4%B8%8A%E4%B8%8B%E6%96%87%E5%8E%8B%E7%BC%A9/</link><description>Recent content in 上下文压缩 on Architecting Intelligence</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Sun, 14 Jun 2026 01:00:00 +0800</lastBuildDate><atom:link href="https://blog.oklensuangoo.uk/zh/tags/%E4%B8%8A%E4%B8%8B%E6%96%87%E5%8E%8B%E7%BC%A9/index.xml" rel="self" type="application/rss+xml"/><item><title>你的 KV Cache，记得你删掉的上下文</title><link>https://blog.oklensuangoo.uk/zh/posts/kv-cache-remembers-deleted-context/</link><pubDate>Sun, 14 Jun 2026 01:00:00 +0800</pubDate><guid>https://blog.oklensuangoo.uk/zh/posts/kv-cache-remembers-deleted-context/</guid><description>免训练地复用 KV cache 的一个子集，往往比从保留文本重算更准。本文给出支配这一现象的变量（coverage）、背后的因果机制（下游注意力痕迹），以及它对 Agent 上下文压缩的直接启示。</description></item></channel></rss>