Splitting agent memory into working and experiential halves lifts tau-bench to 87.9% on Claude Opus 5, a 15.6-point gain

arXiv 2608.24876·high signalXBlueskyLinkedInCopy link

Recuris (arXiv 2608.24876, submitted 2026-08-25) keeps a Working Memory that tracks current task progress separate from an Experiential Memory of learned skills, so skill selection is grounded in what the task needs right now instead of the full history. It improves success in 35 of 37 model-benchmark pairs, gains +17.8 points on GPT-5.6 Sol, reaches +32.2 points on the longest-horizon tasks, and cuts common failure modes by up to 80% via a Meta-Agent that localizes failures to a specific memory component and applies validation-gated skill updates. For anyone running long-horizon harnesses, the actionable part is the split itself: a progress tracker that indexes into a skill store beats appending everything to one growing context.