ERSkill makes memory retrieval itself an evolvable skill, improving agent-memory benchmarks by 31.3%

arXiv·medium signalXBlueskyLinkedInCopy link

ERSkill (arXiv 2608.12720, 2026-08-13) targets a component most agent-memory work leaves static: the retrieval mechanism. It compiles interaction history into a structured store, represents retrieval behaviors as executable skills composed of primitives, and trains a router that matches each query to the skill that builds the right evidence for it. Skill set and router co-evolve during training, with an experience trie recording explored retrieval paths and a double-frontier mechanism decoupling exploration of new skills from the stable set the router deploys against. Reported gains on the average of F1, BLEU-1 and LLM-judge across agent memory benchmarks are 31.3% with Qwen3-Next-80B-A3B-Instruct and 28.1% with GPT-5.4-nano.