Berkeley-Stanford preprint: agents that periodically forget beat ones that remember everything

ombre-ai · reddit · 2026-10-22

A preprint from UC Berkeley, Stanford, and Bespoke Labs (arXiv, Oct 14) tested teams of coding agents on four automated research tasks: heuristic programming, kernel optimization, and model training.
More time, more agents, and more communication all plateaued as history piled up — each agent anchored to its own past attempts and made fewer, smaller edits.
The fix: periodic forgetting — discard every agent's session history and restart from the best result so far. Reported up to 15% better than parallel agents without it, and up to 39% better than a single agent.
Caveats: preprint, bounded research tasks, transfer to everyday agent work unproven. Drew Breunig's related term "context inertia" is one writer's framing, not an established diagnosis.

More from coding & agent