Self-Improvement Memory Enables Continual Recursive Self-Improvement
Self-Improvement Memory Enables Continual Recursive Self-Improvement
Self-Improvement Memory (SIM) enhances performance in Continual Recursive Self-Improvement (CRSI) by recording optimization history. In six cross-dataset comparisons across four domains and four Opus-to-Gemini comparisons, complete SIM scored higher than continuation methods using only harness or root-cause. In nine out of ten independent encoding optimizations, SIM won the comparison, averaging an improvement of 2.36 points (p=0.003). This setup requires a compatible execution interface.
Author affiliations (as listed in the paper): Google Cloud AI · Google DeepMind · Google Cloud AI Research · The Pennsylvania State University.
We started with a practical question: after spending many rounds improving an agent’s harness, what should we carry forward when the model or benchmark changes?
The code is an obvious answer. But it often leaves out why a check was added, what a prompt change was meant to fix, or which earlier attempts failed. That history can matter when the next optimizer has to decide what to keep and what to revise.
We study Continual Recursive Self-Improvement (CRSI) and introduce Self-Improvement Memory (SIM). SIM groups harness edits by the problems they address and records their rationale, predicted and observed outcomes, and evaluation context. The components proposing and selecting edits both read this history. New attempts extend it as optimization continues, while the task-solving model’s weights stay frozen.
Our comparisons start from the same source-evolved harness, with matched target tasks, policy, optimization rounds, and candidate budgets. Across four domains, full SIM achieves higher final scores than harness-only and root-cause-only continuation in the six cross-dataset and four Opus-to-Gemini comparisons we test.
We also run ten independent coding source optimizations, each followed by paired continuations with and without SIM. SIM finishes ahead in nine pairs, with a mean gain of 2.36 score points (95% CI: 1.05–3.68; p = 0.003).
The setting requires compatible execution interfaces. Most main cross-domain comparisons use one run per condition, and matching rounds does not imply equal token or runtime cost. These results support carrying optimization history into continued search; they do not guarantee improvement on every transfer.
The images below show the idea and the coding results, including the negative pair. Happy to discuss the setup or hear how others preserve this kind of experience.
Source: Hugging Face Daily Papers · huggingface.co

