Researchers introduce Recuris, a recursive Experiential-Working Memory architecture designed to address the challenges of recursive self-improvement in long-horizon tasks. The system uses Working Memory to track task progress and guide skill selection from Experiential Memory, grounding decisions in current needs rather than full history.

  • A fixed Meta-Agent processes execution evidence into localized updates for Skill Memory, forming a bounded recursive memory-evolution loop.
  • Recuris improves task success in 35 of 37 completed model-benchmark pairs across four long-horizon benchmarks and ten models.
  • The architecture carries frontier models to SOTA-level performance, adding +17.8 points to GPT-5.6 Sol and +15.6 to Claude Opus 5 on tau-bench.
  • On Qwen3.6-27B/35B, Recuris adds +16.6/+13.5 points on SkillFlow, with advantages widening to +32.2 points on the longest tasks.
  • Common long-horizon failures are reduced by up to 80% as interaction horizons grow.

These results position recursively evolving memory as a scalable foundation for RSI, enabling agents to continuously transform accumulated experience into increasingly effective long-horizon behavior.