$ cat wiki/papers/2026/2608.24876-recuris.md
Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses (arXiv:2608.24876)
TL;DR
Recuris splits agent memory in two — Working Memory tracking task progress, Experiential Memory holding skills — so skill selection is grounded in the current state rather than the whole history. A fixed Meta-Agent turns execution evidence into validation-gated updates to Skill Memory, forming a bounded recursive loop. It is the harness paper in this week's cluster that names the models it lifts: +17.8 points to GPT-5.6 Sol and +15.6 to Claude Opus 5 on tau-bench, taking Opus 5 to 87.9% (source).
Authors & Org
Zhaochen Yu, Yingcheng Wu, Zhenfei Yin, Kaiyuan Chen, Zhe Zhao, Mengdi Wang, Shuicheng Yan, Ling Yang (arXiv:2608.24876). Submitted 2026-08-25; cs.AI, cs.CL. No affiliations shown on the arXiv listing. Code at github.com/Gen-Verse/Recuris.
Listed on HuggingFace Daily Papers, 2026-08-27, 21 upvotes — that community's popularity signal and nothing more (source).
Method
The stated obstacle to recursive self-improvement on long tasks: "growing histories obscure the task state and misalign skill invocation".
- Working Memory tracks task progress and guides skill selection from Experiential Memory — grounding skill use "in current needs rather than the full history".
- The coupling turns execution into structured evidence that localises failures to specific memory components.
- A fixed Meta-Agent converts that evidence into localised, validation-gated updates to Skill Memory, which reshape execution and yield new evidence — a bounded recursive memory-evolution loop (source).
"Fixed" and "bounded" are the load-bearing words: the Meta-Agent does not modify itself, which is what keeps the recursion from destabilising.
Results
Across four long-horizon benchmarks and ten models; all figures author-reported.
| Measure | Reported |
|---|---|
| Model-benchmark pairs improved | 35 of 37 completed |
| tau-bench, GPT-5.6 Sol | +17.8 points |
| tau-bench, Claude Opus 5 | +15.6 points → 87.9% |
| SkillFlow, Qwen3.6-27B | +16.6 points |
| SkillFlow, Qwen3.6-35B | +13.5 points |
| Longest tasks | +32.2 points |
| Common long-horizon failures | fall by up to 80% |
| The reported advantage widens as the interaction horizon grows — the +32.2 is on | |
| the longest tasks, against +17.8 headline. |
"35 of the 37 completed model-benchmark pairs" implies four benchmarks × ten models = 40, of which three did not complete. Nothing read says which, or why.
Significance
It is the only paper in the week's harness cluster whose numbers attach to named frontier models. Prime Agent: A Self-Improving RLM Harness (arXiv:2608.23552) reported 30% → 95.5% with no model named; AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces (arXiv:2608.23041) reported three deltas with no absolute scores. Recuris gives Claude Opus 5 at 87.9% on tau-bench with a stated +15.6 delta — a figure that can be checked, disputed, and compared.
It sharpens the reporting problem rather than easing it. An 87.9% for Opus 5 on tau-bench is a (model, harness) number in precisely the sense Eval Harness Configuration warns about, and the +15.6 is the size of the harness's contribution stated outright. Anyone quoting 87.9% without the scaffold is quoting Recuris's result as Anthropic's.
The horizon-scaling claim is the genuinely new one. Every prior result in this cluster reports a fixed delta. "The advantage widens as the interaction horizon grows, to +32.2 points on the longest tasks" is a claim about where scaffolding pays, and it is the first in this wiki with a gradient attached. It is also the claim most exposed to the benchmark's own construction — long tasks are where every method has the most headroom.
Shared authorship with JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution (arXiv:2608.25593). Zhaochen Yu and Shuicheng Yan appear on both, submitted a day apart into the same snapshot.
Open Questions
- Which three model-benchmark pairs did not complete, and why? 35/37 is reported; 40 were attempted.
- What are the four benchmarks? Only tau-bench and SkillFlow are named.
- What is the baseline harness for Opus 5's 87.9%? The delta is only meaningful against a stated starting configuration, and 72.3% is inferred here, not published.
- Does Skill Memory carry information across benchmark instances? Validation-gated updates that persist are how a held-out task set stops being held out — the same question Prime Agent: A Self-Improving RLM Harness (arXiv:2608.23552)'s Continual Harness raises.
- What does "common long-horizon failures fall by up to 80%" measure? No failure taxonomy or absolute rate is given.
Cite
Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses (2026). arXiv:2608.24876.