Recuris: When AI Learns to 'Refine Experience into Wisdom'
> *"Memory is the soul's alchemist, turning trivial lead into golden wisdom."*
This forum post is an in-depth interpretation of the paper "Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses" (arXiv:2608.24876, Yu et al.). Below is an English rendering of its key content.
The Problem: AI 'Amnesia' in Long-Horizon Tasks
The author opens with an analogy: a cook juggling a stew, onions, and oven bread gets interrupted by the doorbell—and returns to burnt onions and no memory of whether salt was added. Human working memory is famously limited (Miller's 1956 "7±2" chunks). AI agents face the same bottleneck scaled a thousandfold: in tasks lasting hours—browsing, form-filling, price comparison, handling pop-ups—task state gets buried under an ocean of interaction history, and skill invocation becomes unreliable.
Recursive self-improvement (RSI) hits a ceiling here: as history explodes, agents call the wrong skills at the wrong moments.
The Core Idea: Separating Working and Experiential Memory
Borrowing from cognitive neuroscience's distinction between working memory (fast, small, like RAM) and long-term memory (large, slower, like a disk), the paper proposes Recuris, a three-component architecture:
| Human memory | Recuris component | Function | |---|---|---| | Working memory | Task Working Memory | Tracks task progress, guides skill selection | | Long-term/experiential memory | Experiential Memory | Stores organized interaction and skill-execution records | | Procedural memory | Skill Memory | Holds an evolving set of reusable skills |
- Working Memory is like a sticky note in a library: it holds only the minimal key facts needed now (current step, gathered info, missing info)—not the full history. Skill use is anchored to *current* needs rather than the complete context.
- Experiential Memory is the library catalog: a structured archive of all past interactions, selectively activated when working memory needs relevant references.
- Skill Memory evolves dynamically from execution, like internalized bike-riding skills.
- Improvements in 35 of 37 completed model-benchmark pairs
- GPT-5.6 Sol on tau-bench: +17.8 points
- Claude Opus 5 on tau-bench: +15.6 points, reaching 87.9% success
- Qwen3.6-27B on SkillFlow: +16.6 points; Qwen3.6-35B: +13.5 points
- Gains widen with task length: up to +32.2 points on the longest-task subset
- Common long-horizon failure types drop by up to 80%
- Validation cost: every skill update requires testing—an open computational expense.
- Task dependency: experiments focus on tasks with clear success criteria; open-ended tasks are harder to evaluate.
- Memory growth: experiential memory may still bloat over time; principled "forgetting" remains unsolved.
- Yu, Z., Wu, Y., Yin, Z., Chen, K., Zhao, Z., Wang, M., Yan, S., & Yang, L. (2026). *Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses*. arXiv:2608.24876.
- Miller, G. A. (1956). The magical number seven, plus or minus two. *Psychological Review*, 63(2), 81–97.
- Baddeley, A. D. (2000). The episodic buffer: a new component of working memory? *Trends in Cognitive Sciences*, 4(11), 417–423.
The Recursive Evolution Loop
1. Execution produces evidence: each run yields structured evidence of which steps succeeded or failed, and in which memory component. 2. Failure localization: a fixed Meta-Agent pinpoints the specific cause of failure. 3. Validation-gated updates: candidate skill updates are applied only after passing validation tests—"bounded recursion" instead of uncontrolled self-modification. 4. Recursive closed loop: updated skills produce new executions, new evidence, and further refinement.
Experimental Results
Evaluated on four long-horizon benchmarks, including tau-bench (airline customer-service simulation) and SkillFlow (multi-step web automation):
Why It Works
The author identifies three principles behind Recuris's effectiveness:
1. Separation of concerns — each memory component optimizes independently (what's needed now / what happened before / how to do better). 2. Bounded recursion — validation-gated updates make RSI safe and predictable. 3. Learning from execution — the agent learns from real outcomes, including failures, rather than only demonstration data.
The design also shows convergent evolution with Baddeley's human working-memory model (episodic buffer, central executive), suggesting effective intelligence may require shared architectural principles regardless of substrate.
Acknowledged Limitations
Closing Thought
> "These results position recursively evolved memory as a scalable foundation for RSI, enabling agents to continuously convert accumulated experience into increasingly effective long-horizon behavior."
Learning, the author concludes, is not about remembering more—it's about knowing when to use what. Recuris gives agents, for the first time, the ability to "refine" raw experience into effective action.