New ByteDance paper shows for test-time improvement, keeping the messy history of attempts can work better than turning that history into a neat memory summary.

Chain-of-Experience keeps earlier attempts and feedback in context, then asks the model to try again. Across 6 math, coding, and knowledge benchmarks, self-feedback averaged 71.0%, versus 66.8% with iterative solving but no feedback; correctness or executor feedback reached 79.3%.

The paper also reports 5.6% overall improvement with 19% lower API cost across tasks and models when feedback is used.

No weights change here, so this is contextual adaptation rather than persistent learning. Self-feedback also hurt on BrowseComp-Plus when solving the task required external search.

For agents, preserve the trajectory, add reliable feedback, and compress only when you know what can safely disappear.