面向长时域智能体框架的递归经验-工作记忆演化机制

摘要

递归自我改进（RSI）在长时程任务中仍然困难，因为不断累积的历史记录会模糊任务状态，并导致技能调用错位。我们提出 Recuris，一种用于长时程智能体框架的递归经验-工作记忆架构，其中工作记忆跟踪任务进度并引导从经验记忆中进行技能选择，使技能使用基于当前需求而非完整历史。这种耦合还将执行过程转化为结构化证据，从而将失败定位到具体记忆组件。跨任务地，一个固定的元智能体将这些证据转化为对技能记忆的局部化、验证门控更新，重塑执行行为并产生新证据，形成一个有界的递归记忆演化循环。在四个长时程基准和十个模型上，Recuris 在 37 个完成的模型-基准组合中改善了 35 个的任务成功率，将前沿模型提升至 SOTA 级任务成功率：在 tau-bench 上，为 GPT-5.6 Sol 增加 17.8 个百分点，为 Claude Opus 5 增加 15.6 个百分点，使 Opus 5 达到 87.9%；在 SkillFlow 上，为 Qwen3.6-27B/35B 分别增加 16.6/13.5 个百分点。随着交互时程增长，优势进一步扩大，在最长的任务上达到 +32.2 个百分点，常见的长时程失败减少多达 80%。这些结果将递归演化的记忆定位为 RSI 的可扩展基础，使智能体能够不断将累积经验转化为日益高效的长时程行为。代码：https://github.com/Gen-Verse/Recuris

English

Recursive self-improvement (RSI) remains hard in long-horizon tasks, where growing histories obscure the task state and misalign skill invocation. We introduce Recuris, a recursive Experiential-Working Memory architecture for long-horizon agent harnesses, in which Working Memory tracks task progress and guides skill selection from Experiential Memory, grounding skill use in current needs rather than the full history. This coupling also turns execution into structured evidence that localizes failures to specific memory components. Across tasks, a fixed Meta-Agent turns that evidence into localized, validation-gated updates to Skill Memory that reshape execution and yield new evidence, forming a bounded recursive memory-evolution loop. Across four long-horizon benchmarks and ten models, Recuris improves task success in 35 of the 37 completed model-benchmark pairs, carrying frontier models to SOTA-level task success: on tau-bench it adds +17.8 points to GPT-5.6 Sol and +15.6 to Claude Opus 5, taking Opus 5 to 87.9%, and +16.6/+13.5 points on Qwen3.6-27B/35B on SkillFlow. The advantage widens as the interaction horizon grows, to +32.2 points on the longest tasks, and common long-horizon failures fall by up to 80%. 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. Code: https://github.com/Gen-Verse/Recuris