Google paper proposes Self-Improvement Memory for harness optimization
The non-obvious implication is that naive harness evolution breaks when the underlying model or benchmark shifts, and structured memory of edit rationales is the fix. Developers building automated agent pipelines or self-improving coding systems can apply the SIM pattern directly: log predicted versus observed outcomes for every change. Teams comparing Claude Opus and Gemini setups get concrete transfer-cost numbers, and the paper includes a Chat with Paper resource from dair.ai for quick exploration.