Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses introduces Recuris, an architecture designed to improve how AI agents handle long-horizon tasks. The research addresses the tendency of agents to lose track of goals and misalign skill usage as interaction histories grow. By decoupling memory into a structured Working Memory (WM) and an Experiential Memory (EM), the system grounds skill selection in current task needs rather than the entire conversation history, enabling agents to recursively improve their performance through a validation-gated loop.
The Problem with Long-Horizon Tasks
In long-horizon tasks, agents often struggle because their memory mechanisms are fixed while the task requirements evolve. Existing methods typically retrieve information from the full interaction history or initial instructions, which becomes unreliable as the context fills with completed steps, noise, and outdated information. This leads to agents invoking skills that are no longer relevant to the current state of the task. The authors argue that the core issue is not a lack of experience, but the absence of a reliable, compact representation of the current task state to guide that experience.
How Recuris Works
Recuris introduces a memory-control layer that operates through two primary loops:
Within-Task Coupling: The system uses a Working Memory (WM) to track progress and unresolved goals. This WM acts as an interface, ensuring that when the agent needs to invoke a skill from Experiential Memory (EM), the selection is based on the current verified state of the task. After an action is taken, a checker evaluates the environment's response to update the WM, creating a closed loop of state-grounded execution.
Cross-Task Evolution: The system records structured execution traces that link actions, observations, and state updates. A fixed Meta-Agent analyzes these traces to localize failures to specific memory components (EM, WM, the invocation policy, or the checker). It then proposes targeted, component-specific patches. These patches are only admitted if they pass a validation gate on held-out tasks, ensuring that the agent does not regress on existing abilities.
Performance and Results
The authors evaluated Recuris across four long-horizon benchmarks and ten models, ranging from 3B open-weight models to frontier models. The system improved task success in 35 of 37 completed model-benchmark pairs. Key results include:
Frontier Model Gains: On the τ²-Bench, Recuris added 17.8 points to GPT-5.6 Sol and 15.6 points to Claude Opus 5, bringing the latter to an 87.9% success rate.
Horizon Scaling: The performance advantage increased as tasks grew longer, reaching a 32.2-point improvement on the longest tasks.
Failure Reduction: Common long-horizon failure modes were reduced by up to 80%.
Localization Accuracy: The structured traces allowed for significantly more accurate failure localization (64.8%) compared to relying on final task outcomes alone (13.0%).
Key Considerations
The architecture maintains the underlying LLM and the outer agent harness as fixed components. Recursion is confined to the externalized memory-control layer, meaning the base model itself is not retrained or modified. This approach allows for continuous improvement of the agent's behavior through the evolution of its memory mechanisms, rather than requiring changes to the model's weights. The authors note that because the Meta-Agent and validation procedures are fixed, the system provides a bounded, stable framework for recursive self-improvement.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!