Artificial intelligence systems often struggle with tasks that require planning and decision-making over long periods. As the history of actions and events grows, it becomes harder for AI to remember what matters and choose the right skills at the right time. A new research paper introduces a novel memory architecture called Recuris that helps AI agents better manage and learn from their experiences, significantly improving their ability to handle complex, long-term tasks. This advancement could pave the way for smarter AI that adapts and improves continuously as it works on challenging problems.
Key Takeaways
- Recuris architecture: Combines two types of memory—Working Memory to track current task progress, and Experiential Memory to store past skills and experiences—enabling more focused and relevant skill selection.
- Recursive learning loop: The system uses feedback from task execution to identify where mistakes happen, then updates its skill memory in a targeted way, improving future performance.
- Strong performance gains: Tested on four long-horizon benchmarks and ten AI models, Recuris improved task success in 35 out of 37 cases, even boosting state-of-the-art models like GPT-5.6 Sol and Claude Opus 5 by over 15 percentage points.
- Better with longer tasks: The benefits of Recuris increase as tasks grow longer and more complex, reducing common failure types by up to 80% on the toughest challenges.
Long-horizon tasks pose a unique challenge for AI because the agent must remember and reason over a lengthy sequence of steps. Traditional AI systems often keep a single, growing record of everything that happened, which can become overwhelming and confusing. This can lead to poor decisions because the AI struggles to figure out which past experiences are relevant to the current moment.
Recuris addresses this by introducing two distinct but connected memory systems. The Working Memory acts like a dynamic to-do list or progress tracker, keeping tabs on what the agent has done so far and what it needs to do next. Meanwhile, the Experiential Memory stores skills and knowledge gained from previous experiences. Instead of sifting through the entire history of actions, the AI uses the Working Memory to focus on which skills from the Experiential Memory are most useful right now. This approach helps ground skill selection in the current context, avoiding confusion caused by irrelevant past details.
Another key innovation is the recursive feedback loop. When the AI executes skills, it generates evidence about how well those skills worked. The system can then pinpoint which part of its memory led to mistakes—whether it was in tracking progress or choosing skills. A Meta-Agent uses this information to selectively update the Skill Memory, refining the AI’s abilities over time. This creates a bounded cycle of memory evolution, where the AI continually learns from its successes and failures in a structured way.
The researchers tested Recuris on multiple challenging benchmarks that require long-term planning, using a variety of advanced AI models. The results showed consistent improvements, with some models achieving state-of-the-art success rates. Notably, the improvements were even more pronounced for longer tasks, demonstrating that Recuris scales well as the complexity of the problem increases.
Looking ahead, this research suggests that recursively evolving memory architectures like Recuris could form a scalable foundation for AI systems capable of recursive self-improvement. By continuously transforming accumulated experience into better skills and strategies, AI agents might handle increasingly complex tasks without human intervention. While more work is needed to explore real-world applications, this new approach marks a promising step toward more adaptive and capable long-horizon AI.
Based on research published on arXiv by Zhaochen Yu, Yingcheng Wu, Zhenfei Yin et al..