Sakana AI Researcher: Memory Harnesses Critical for Long-Running Agents, But Strategy Matters

AI Engineer · youtube · 2026-08-12

In a talk, Stefania Druga from Sakana AI shared experiments on memory harnesses for long-running research agents. She found that when all info fits in context, adding memory doesn't help and increases cost; but for long-horizon tasks, it's crucial. She proposed a write-manage-read control loop and compared recall policies: none, vector RAG, decisions ledger, and oracle. On 68 xbench questions, the ranked ledger performed best, even beating the oracle, because giving correct memory doesn't ensure use. Bad memory is expensive. She ran experiments on a local M3 Ultra.

Original post →

More from AGI Musings

AGI Musings channel →