Rohan Paul (@rohanpaul_ai)T2·Agents
78 pts
A new Google and Penn State paper proposes Self-Improvement Memory (SIM): when an LLM rewrites an agent's harness, the code that runs prompts and tools, each edit is logged with the problem it targeted, its reasoning and its measured result. After switching models or benchmarks, an optimizer reads this log to decide which prior edits to keep or revise, preventing regressions. Moving a Claude Opus 5.5 harness to Gemini 3.8 Flash workspace tasks, SIM raised scores from 68.41 to 84.75 versus 73.42 without the log, and led in 9 of 10 repeated coding runs by 2.36 points on average.