In large language model (LLM) reasoning, thought chains are ephemeral: they vanish with the context window, pruned search branches leave no record, and memory buffers cannot be diffed, merged, or audited. Unlike other complex software processes (code, infrastructure, data, experiments), reasoning is not version-controlled. We introduce GitOfThoughts, which stores an agent's reasoning tree as a git repository: every scored thought is a commit, scores are notes, outcomes are tags, and retrieval is a 'git log' over the agent's own history. This makes reasoning replayable, auditable, and mergeable across agents at near-zero engineering cost.
We then tackle the more complex question: does memory, in any substrate, actually improve accuracy? Across five substrates (none, markdown, vector, graph, git), two benchmarks, two model scales, and pre-registered replications, the answer for novel problems is no. No memory format reliably helps, and a promising early result collapsed under its own pre-registered replication. Memory pays only above what we call the copyability threshold: when the retrieved case is a near-duplicate of the current problem (similarity ~ 0.8), accuracy jumps sharply; below it, nothing. The gain is answer retrieval, not method transfer: a 4.5x larger model doubles the near-duplicate payoff yet still cannot extract a transferable method from a worked example. The only general lever we find is test-time sampling. Therefore, the case for git-as-substrate is auditability, provenance, and mergeability at accuracy parity. We document a retracted result and a refuted hypothesis to model the evaluation standard we hold ourselves to.
Blogger's Review: GitOfThoughts presents an innovative approach to managing an agent's reasoning processes through version control, enabling auditable and replayable thought chains. However, the effectiveness of memory appears limited, highlighting the complex relationship between data similarity and accuracy in agent learning.