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

GitOfThoughts introduces a novel framework for version-controlling reasoning in large language models (LLMs) by storing an agent's reasoning tree as a git repository, allowing for replayable, auditable, and mergeable reasoning processes. The study evaluates the impact of memory across five different substrates (none, markdown, vector, graph, git) and finds that memory does not reliably improve accuracy for novel problems, with significant gains only occurring when retrieved cases are near-duplicates of current problems. This framework emphasizes the importance of auditability and provenance in LLM reasoning, making it a valuable tool for practitioners focused on reproducibility and traceability in AI systems.