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

In one reported experiment, extra memory added nothing when the research task fit within what the model could already read. The talk description reports unchanged accuracy and higher cost when every paper fit within the model's context window, the material available to it at once.

For longer tasks, the experiment held the model fixed and changed how earlier information was retrieved. A ledger that recorded and prioritized decisions was the best-performing recall policy reported across 68 questions. Even supplying the correct memory directly did not yield perfect performance. The useful distinction is between retrieving relevant information and getting the model to use it correctly.

“On a literature review task where every paper already fit inside the context window, adding a memory harness changed nothing: the same accuracy, at higher cost.”

AI Engineer

“Move to a long horizon task where the answer sits at step 124 and the question arrives at step 500, far outside the window, and the harness becomes the entire game.”

AI Engineer

“Her framing is that memory is a write, manage, read control loop wrapped around the model, not a database you attach to it.”

AI Engineer

“The oracle pointedly does not reach the ceiling, because giving a model the right memory does not make it use the right memory.”

AI Engineer

“Ranked recall was also cheaper, which is the line worth keeping: bad memory is expensive, since it burns tokens and sends the agent the wrong way.”

AI Engineer