Report
Researchers report AI harness gains from remembering why the code changed
An arXiv paper reports that remembering past code edits helps harness optimizers adapt across models and benchmarks.
TLDR
The authors’ Self-Improvement Memory logs why an AI agent’s harness—the code running prompts and tools—was edited, plus expected and observed results. In their arXiv paper, they report an average 2.36-point final-score gain over harness-only continuation across ten coding runs. This is memory for the harness optimizer, not the task-solving model: keep the reasoning, not just the code.
Combined views
26.5K
3 Sources, first seen ago

Comments