Article Proposes Ranking Long-Context Agent Memory by Value Instead of Similarity

On November 4, 2026, at 08:19, Juejin published an article titled "Agent Memory Should Be Addressed by Value, Not Similarity" (original Chinese title: Agent 的记忆该按价值排地址而不是按相似度). The piece examines the pain points that long-context Agents encounter during repeated retrieval. Its central argument is that an Agent's memory retrieval should be ordered by value rather than by similarity, a reordering intended to improve the efficiency of attention mechanisms and RAG. The article observes that at every step, a long-context Agent repeatedly tackles highly similar retrieval problems, but attention and RAG each... (the quoted passage breaks off at this point, so the full argument cannot be verified from the available material). This is the only report found so far, and no follow-up developments have been recorded.