Memory Harnesses for Long-Running Research Agents — Stefania Druga, Sakana.ai
Stefania Druga from Sakana.ai tested how AI agents can remember information during extended tasks.
She built a memory system that functions as a write-manage-read loop around the model, rather than a simple database. The key finding was negative: when all information fit within the model's context window, memory made no difference—only increased cost. However, when the answer lay far ahead—at step 124 while the question appeared at step 500—memory became critical. Her best approach was a ranked list of previous decisions, which outperformed both random search and an "oracle" with perfect memory. She ran the experiments on a local computer in Tokyo with fans around it.
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
