--- title: How LLMs Decide What to Forget: KV Cache Eviction Explained description: When GPU memory fills during long-context LLM inference, the KV cache must evict tokens. Transformer attention is naturally sparse — only 10–20% of tokens... canonical: https://daily.dev/posts/how-llms-decide-what-to-forget-kv-cache-eviction-explained-k6kfg62tc twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: How LLMs Decide What to Forget: KV Cache Eviction Explained | daily.dev og:description: When GPU memory fills during long-context LLM inference, the KV cache must evict tokens. Transformer attention is naturally sparse — only 10–20% of tokens... og:url: https://daily.dev/posts/how-llms-decide-what-to-forget-kv-cache-eviction-explained-k6kfg62tc og:image: https://api.daily.dev/og/posts/K6KFg62tc.png og:image:alt: How LLMs Decide What to Forget: KV Cache Eviction Explained og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # How LLMs Decide What to Forget: KV Cache Eviction Explained **[Medium](https://daily.dev/sources/medium_js)** · 3 min read · 2 upvotes · 0 comments ## Summary When GPU memory fills during long-context LLM inference, the KV cache must evict tokens. Transformer attention is naturally sparse — only 10–20% of tokens accumulate most attention scores — making selective eviction viable. H2O (Heavy Hitter Oracle) tracks cumulative attention scores and drops low-scoring tokens while preserving recent ones, achieving 2–5x compression with ~1–2% quality loss, but works reactively during decoding. SnapKV targets a different problem: near-duplicate KV vectors in RAG workloads with overlapping retrieved chunks. It clusters key vectors via k-means and keeps one representative per cluster, achieving 4–8x reduction with 1–3% quality loss, especially effective on repetitive document types. The two algorithms are orthogonal — H2O compresses by attention weight, SnapKV by content redundancy — and production systems typically layer both on top of LRU-based cross-request eviction. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Similar posts on daily.dev - [Why An LLM’s Memory Gets Expensive and How to Fix It](https://daily.dev/posts/why-an-llm-s-memory-gets-expensive-and-how-to-fix-it-tfmmfttjf) · ByteByteGo · 2 upvotes · 0 comments - [Rethinking KV Caching For Production Inference](https://daily.dev/posts/rethinking-kv-caching-for-production-inference-1px3uv98r) · Daily Dose of Data Science \| Avi Chawla \| Substack · 0 upvotes · 0 comments - [How KV Caching Slashes LLM Inference Costs at Scale](https://daily.dev/posts/how-kv-caching-slashes-llm-inference-costs-at-scale-7g9k0lnod) · DigitalOcean Community · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#rag](https://daily.dev/tags/rag), [#ai-inference](https://daily.dev/tags/ai-inference) [View this post on daily.dev](https://daily.dev/posts/how-llms-decide-what-to-forget-kv-cache-eviction-explained-k6kfg62tc) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"How LLMs Decide What to Forget: KV Cache Eviction Explained","url":"https://daily.dev/posts/how-llms-decide-what-to-forget-kv-cache-eviction-explained-k6kfg62tc","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/how-llms-decide-what-to-forget-kv-cache-eviction-explained-k6kfg62tc"},"datePublished":"2026-07-24T05:11:09.824Z","dateModified":"2026-07-24T05:12:36.535Z","description":"When GPU memory fills during long-context LLM inference, the KV cache must evict tokens. Transformer attention is naturally sparse — only 10–20% of tokens...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/42ab9c273fc2966e64847b0808a81c6b?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/42ab9c273fc2966e64847b0808a81c6b?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Medium","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Medium","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/medium","url":"https://daily.dev/sources/medium_js"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/how-llms-decide-what-to-forget-kv-cache-eviction-explained-k6kfg62tc","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,rag,ai-inference","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Medium","item":"https://daily.dev/sources/medium_js"},{"@type":"ListItem","position":3,"name":"How LLMs Decide What to Forget: KV Cache Eviction Explained"}]} ```