HORMA: Hierarchical 'Organize-then-Retrieve' Working Memory for Long-Horizon Agents

arXiv·low signal

A new arXiv paper (2606.11680) proposes HORMA, which attacks the statelessness that degrades agent reasoning and inflates inference cost on long tasks by splitting working memory into structured memory construction plus navigation-based retrieval over a hierarchical experience store. Rather than dumping raw history into context, the agent organizes experience and then navigates to the relevant slice. It's part of a visible cluster of June agent-memory work and is directly relevant to anyone building durable, multi-session coding or research agents.