Can an agent recursively improve—without updating a single model weight? Today, we introduce Recuris: Recursive Experiential–Working Memory Evolution for Long-Horizon Agent Harnesses. Recuris enables recursive self-improvement through two tightly coupled loops around a frozen LLM: 1. Within-task execution: Verified EM–WM Coupling Working Memory maintains a compact, evidence-grounded representation of the agent’s current progress and unresolved goals. This evolving task state determines when—and which—reusable skill should be invoked from Experiential Memory. After the agent acts, environment feedback verifies whether progress actually occurred before Working Memory is updated. The result is a closed execution loop: Task State → Skill Invocation → Execution → Verification → Updated Task State 2. Across-task improvement: Recursive Skill Memory Evolution EM–WM coupling turns execution into structured diagnostic evidence. A fixed Meta-Agent analyzes these traces, localizes failures to specific memory components—such as the stored skill, working-state representation, invocation policy, or checker—and proposes targeted repairs. Candidate updates are admitted only after passing a fixed validation gate. Once admitted, the evolved memory changes future execution, producing new behavior and new evidence for the next round of improvement: Memory → Behavior → Evidence → Failure Localization → Targeted Update → Better Memory This is the core recursive loop of Recuris: memory shapes behavior, behavior reveals the limitations of memory, and those limitations guide the next memory update. Across 4 long-horizon benchmarks and 10 models, Recuris improves task success in 35 of the 37 completed model–benchmark pairs, spanning open-weight models as small as 3B parameters and today’s strongest frontier models. On τ²-Retail: • GPT-5.6 Sol: 58.3 → 76.1, +17.8 points • Claude Opus 5: 72.4 → 87.9, +15.6 points Recuris takes Claude Opus 5 to 87.9% task success—9.7 points above the best result achieved by any model in our evaluation without Recuris. The advantage becomes larger as the interaction horizon grows, reaching +32.2 points on the longest tasks, while common long-horizon failure modes are reduced by up to 80%. Crucially, evolved memory transfers across models. A memory evolved on a separate deployment model can be loaded unchanged into GPT-5.6 Sol and Claude Opus 5—even though neither model participated in its evolution. These results suggest a practical and scalable path toward recursive self-improvement: An agent does not need to begin by rewriting its own weights. It can first evolve the external memory mechanisms that determine what it remembers, what remains unresolved, when experience should be invoked, and how failures reshape future execution. The model remains frozen. The memory keeps evolving. The agent becomes increasingly capable. 📄 Paper: arxiv.org/abs/2608.24876 💻 Code: github.com/Gen-Verse/Recuris