Compile recurring intent into specialist agents instead of retrieving memories: 88.9% win rate over a memory-augmented assistant

arXiv 2608.08995·medium signalXBlueskyLinkedInCopy link

Muscle Memory (arXiv 2608.08995, Aug 10) argues that the dominant memory pattern — store experience as text, embeddings, reflections, or rules and let a general orchestrator interpret it at inference time — is the wrong default for personalization, because it leaves users paying a multi-turn tax re-correcting format, depth, and scope. Instead it compiles recurring user intent into purpose-built specialist agents via a four-phase pipeline: harvest patterns from conversation history, analyze to separate behavioral from task patterns, augment the candidate specialists, and quality-gate them with two-stage trigger matching. Across 90 held-out scenarios and five personas, the specialist won 32 of 36 cases where one activated (88.9%), for +2.05 personalization on a 1–4 scale at a −0.28 accuracy cost.