GitOfThoughts: Version-Controlled Reasoning and Agent Memory You Can Replay, Diff, and Merge

In three linesGitOfThoughts stores LLM agent reasoning as a git repository: each thought is a commit, scores are notes, outcomes are tags. Empirical study across 5 memory substrates (none, markdown, vector, graph, git): memory improves accuracy only when retrieved cases are near-duplicates of current problems (similarity >0.8). Main lever remains test-time sampling.

Summary generated by Claude — human-verified