What happened
A new system called LiveCoder lets AI models remember what went wrong in previous attempts at writing code across an entire software repository, then use that knowledge to do better on the next try. In practice, this means the same AI can now solve harder coding problems by learning from its own mistakes instead of starting from scratch each time.
Why it matters
Until now, when AI tried to write code for a large codebase multiple times, each attempt was isolated — the AI didn't learn from what broke or what worked before, so it kept making the same mistakes or rediscovering the same solutions. LiveCoder changes this by keeping a memory of successful code, failed approaches, and the best version found so far, which means the AI gets smarter with each retry on the same task. The measurement matters: on one standard test, this approach improved success rates by 23 percentage points and cut computational costs in half. What becomes possible is using cheaper, slower AI models repeatedly with memory instead of paying for bigger, faster models to solve it once.