[research] Recuris: Working Memory + Skill Selection Cuts Long-Horizon Agent Failures

[github-actions[bot]]

🔬 The Finding

Researchers introduced Recuris, a recursive Experiential-Working Memory architecture for long-horizon agents. Instead of reasoning over full history (which degrades skill selection), a Working Memory component tracks current task state and selects skills from Experiential Memory grounded in present needs. A Meta-Agent converts execution failures into localized, validation-gated memory updates — creating a bounded recursive self-improvement loop. Evaluated across four long-horizon benchmarks and ten models.

⚙️ What It Means for Agentic Workflows

1. Replace full-history prompting with structured working memory: agents that track task state explicitly — not raw context windows — make better tool/skill selections and fail more predictably.
2. Failure localization is a design primitive: when your agent errs, the architecture pinpoints which memory layer broke down (skill, experience, working state), enabling targeted fixes rather than full-pipeline reruns.

🔗 Source

Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses — August 26, 2026

Generated by Daily Agentic AI Research Digest · 87.3 AIC · ⌖ 11.4 AIC · ⊞ 24.2K · ◷

• expires on Sep 3, 2026, 9:56 AM UTC