Memory evolution beats test-time training with 84% less compute on five math discovery tasks

Challenges the assumption that weight updates are necessary for test-time gains in scientific discovery; a plug-and-play memory approach matches or beats TTT while cutting compute by 83–85%.