Recuris: Memory as Alchemy — Experiential–Working Memory Evolution for Long-Horizon AI Agents

Forum topic · 小凯 · 2026-08-26

Summary

This forum post is a detailed Chinese-language interpretation of the paper 'Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses' (arXiv:2608.24876), which introduces Recuris, a memory architecture for AI agents tackling long-horizon tasks. The core insight borrows from human cognition: separating working memory (tracking current task state and guiding skill selection) from experiential memory (storing historical interactions) and a skill memory that evolves dynamically. A fixed Meta-Agent analyzes structured execution evidence to localize failures to specific memory components, and updates are applied only after validation-gated testing, forming a safe recursive self-improvement loop. Reported results: Recuris improved performance in 35 of 37 model-benchmark pairs, including +17.8 points on tau-bench for GPT-5.6 Sol, +15.6 points (reaching 87.9%) for Claude Opus 5 on tau-bench, and gains of +13.5 to +16.6 on SkillFlow for Qwen3.6 models. Advantages grew with task length (+32.2 points on the longest tasks), and common long-horizon failure modes dropped by up to 80%. The post also discusses limitations such as validation cost, task-dependence of success criteria, and unbounded experiential memory growth.

> *"Memory is the soul's alchemist, turning trivial lead into golden wisdom."*

This post is a deep-dive interpretation (in Chinese, translated here) of the paper "Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses" (Yu et al., arXiv:2608.24876). It explains why AI agents "forget" during long-horizon tasks and how the proposed Recuris architecture solves this by mimicking human memory systems.

The problem: AI amnesia in long-horizon tasks

The author opens with an analogy: a cook distracted mid-recipe forgets whether salt was added — a limitation of limited working memory (Miller's classic "7±2"). Scaled up, AI agents performing tasks spanning hundreds of steps (browsing, form-filling, price comparison, multi-turn dialogues) suffer from exponential growth of interaction history: critical early information gets drowned in noise, causing wrong skill invocations and wrong actions at wrong times. This blocks Recursive Self-Improvement (RSI), which requires an agent to observe its own failures and improve.

The Recuris architecture: a memory trio

The key insight borrows from cognitive neuroscience — human memory is not a single system:

Task Working Memory — like human working memory (RAM-like): tracks current task progress (e.g., flight search → price comparison → payment), holds only minimal task-relevant information, and guides skill selection from experience memory, anchoring behavior in present needs rather than full history.
Experiential Memory — the structured archive of all historical interactions, selectively activated when working memory needs relevant references.
Skill Memory — not pre-programmed rules but skills that dynamically evolve from execution, analogous to internalized skills like riding a bike.

How skills evolve: a safe recursion loop

1. Execution produces evidence — structured records of which steps succeeded/failed and in which memory component. 2. Failure localization — a fixed Meta-Agent pins down the exact cause of failure. 3. Validation-gated updates — candidate updates are applied only after passing verification tests (addressing the classic "runaway recursion" risk of RSI). 4. Recursive closure — improved skills generate new executions, new evidence, and further refinement.

The paper frames this as three design principles: separation of concerns (each memory component optimizes independently), bounded recursion (safety via validation), and learning from execution rather than from human demonstration data. The author also notes a striking convergence with Baddeley's working memory model (episodic buffer, central executive), suggesting effective intelligence may share common architectural principles.

Experimental results

Evaluated across four long-horizon benchmarks (including tau-bench, an airline customer-service simulation, and SkillFlow, multi-step web automation):

Improvements on 35 of 37 completed model–benchmark pairs.
GPT-5.6 Sol: +17.8 on tau-bench.
Claude Opus 5: +15.6 on tau-bench, reaching 87.9% success.
Qwen3.6-27B: +16.6 and Qwen3.6-35B: +13.5 on SkillFlow.
Gains widen with task length: +32.2 points on the longest-task subset.
Common long-horizon failure types drop by up to 80%.

Limitations

The post honestly covers the paper's caveats:

1. Validation cost — every skill update requires verification, which is computationally expensive. 2. Task dependence — experiments focus on tasks with clear success criteria; open-ended tasks (creative writing, strategic planning) are harder to evaluate. 3. Memory capacity — experiential memory may still grow unboundedly over time; principled "forgetting" remains unsolved.

Conclusion

The paper closes poetically: recursive memory evolution positions itself as a scalable foundation for RSI, letting agents continuously convert accumulated experience into increasingly effective long-horizon behavior. As the author summarizes: learning is not remembering more — it's knowing what to use, and when.

References

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.