Recuris Evolves Memory for Long-Horizon Agents

Recuris makes recursive self-improvement more inspectable by turning memory—not model weights—into the optimization surface. The results are impressive, but the framework’s real test will be whether its gains survive noisier production tasks and model-specific drift.

• –Working Memory tracks verified task state, preventing skill retrieval from being buried in an ever-growing conversation history
• –A fixed meta-agent patches only the implicated memory component, while held-out validation blocks unproven updates
• –Recuris improved 35 of 37 completed model-benchmark pairs, with gains reaching 32.2 points on the longest tasks
• –Model-specific evolution matters: the repository reports that a rebuilt GPT-OSS-20B memory gained 10.2 points, while a general-purpose package transferred negatively
• –The approach offers developers a practical middle ground between static prompting and expensive fine-tuning, with auditable memory artifacts that can be versioned and tested