https://blog.oklensuangoo.uk/tags/llm/Recent content in LLM on Architecting IntelligenceHugoen-usSun, 14 Jun 2026 01:00:00 +0800https://blog.oklensuangoo.uk/posts/kv-cache-remembers-deleted-context/Sun, 14 Jun 2026 01:00:00 +0800https://blog.oklensuangoo.uk/posts/kv-cache-remembers-deleted-context/Reusing a subset of a KV cache — training-free — is often more accurate than recomputing from the kept text. This post gives the variable that governs it (coverage), the causal mechanism behind it (a downstream-attention trace), and what it means for agent context compression.