Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses

Imagine you're building an AI that needs to handle complex, ongoing tasks over extended periods, like managing a project or coordinating multiple actions. The challenge lies in keeping track of what has happened and making decisions based on that history without losing sight of the current context. Traditional approaches often struggle here, leading to issues like forgetting important details, misapplying skills, or failing to adapt when things go wrong — this is what's called context drift or skill misalignment. These failures can severely limit an agent's effectiveness in real-world applications where tasks are not neatly segmented or predictable.

To address these challenges, Recuris introduces a recursive memory architecture that allows an agent to maintain a dynamic understanding of its task environment. By using a Working Memory that tracks progress and an Experiential Memory that informs skill selection based on immediate needs, Recuris enables agents to adaptively refine their actions. This method not only helps in identifying where failures occur but also facilitates targeted updates to the agent's skills based on localized evidence. The results are promising: across various long-horizon benchmarks, Recuris significantly boosts task success rates, even achieving state-of-the-art performance in several cases. For builders, this means a more robust framework for developing agents that can learn and improve over time, making them more reliable in complex, real-world scenarios.

Recuris introduces a novel architecture for managing long-horizon tasks through recursive memory evolution.

The results show significant improvements across multiple benchmarks, though the evaluation could benefit from more diverse scenarios.