We built a learning loop that learns from its own failures

We recently published our research on the learning loop behind the Dropstone agent. Instead of throwing more compute at a problem or trying to dynamically update weights, we tested what happens when you give a frozen open-weight model persistent memory of its own mistakes.

We found that by keeping three stores the Experience, Lessons, and Knowledge the agent successfully builds up scar tissue. In a held-out study, it was able to independently rediscover hidden mathematical laws and coding rules purely from trial, error, and data.