Memory Harnesses for Long-Running Research Agents — Stefania Druga, Sakana.ai

Long-horizon AI agents frequently suffer from context bloat, leading to contradictory outputs and task drift. A robust memory harness, structured as a write-manage-read loop, addresses this by integrating specific recall and archival mechanisms. Experiments using local models like Qwen 27B and DeepSeek V4 Flash reveal that a ranked recall policy significantly outperforms standard retrieval methods on complex benchmarks such as XBench. While memory is unnecessary for tasks fitting within a model's context window, it is critical for long-running operations. Notably, providing accurate ground-truth memory—an "oracle"—does not guarantee success, as models must still effectively interpret and prioritize retrieved information. Establishing a formal recall policy serves as a first-class metric for optimizing agentic performance, reducing token costs, and maintaining operational sovereignty in local AI deployments.