Recuris iterative memory architecture resolves agent failure in long-horizon tasks and cuts execution errors by 80%

A new research paper published by a team of researchers, including Zhaochen Yu, Ling Yang, and Shuiqing Yan, introduces an innovative architecture for AI agents called Recuris, designed to address one of the most complex technical challenges in autonomous agent development: iterative self-improvement and the degradation of model accuracy in long-horizon tasks.

The new architecture relies on a structural separation between two types of memory:working memory, which tracks the task trajectory step by step and determines immediate requirements, andexperiential memory, which stores accumulated skills and experience. This design addresses a common failure mode in conventional agents, where bloated conversation histories and logs of past operations obscure the task state and trigger poor tool or skill retrieval, restricting skill selection to the actual need of the current step rather than flooding the model with the entire operational history.

The system converts execution trajectories into structured evidence that isolates and pinpoints causes of failure within memory components. A fixed meta-agent then translates this evidence into localized updates governed by rigorous verification tests on skill memory. This process yields a controlled, iterative memory evolution loop that reshapes execution and progressively generates fresh evidence without compromising system stability.

Test results across four long-horizon benchmarks and ten different models showed marked gains, with Recuris improving performance in 35 out of 37 completed test cases. On tau-bench, the architecture raised GPT-5.6 Soul accuracy by 17.8 points and added 15.6 points to Claude Opus 5, bringing its success rate to 87.9%. It also recorded gains of 16.6 and 13.5 points on two versions of Qwen 3.6 (27B and 35B) on the SkillFlow benchmark. The performance margin widened in more complex tasks to reach 32.2 points, with common operational failures dropping by up to 80%.

This architectural shift resets calculations for development teams and innovation hubs across the Gulf, Egypt, and the wider region. Organizations building agents to automate administrative workflows, supply chains, or enterprise system queries face a primary obstacle: agent efficiency deteriorates as task steps multiply, while inference costs climb as thousands of tokens are consumed by context history. Moving to an evolving iterative memory model allows mid-sized open-source models, such as Qwen variants, to deliver performance comparable to massive proprietary models, lowering compute infrastructure costs and keeping operational data securely in-house without requiring full weight retraining.

The findings demonstrate that disciplined memory evolution provides a scalable foundation for agent self-improvement, opening the way to turning accumulated operational experience into more mature, consistent execution across complex digital environments.