A research team has mapped AI agent reasoning onto git commits, then turned around and published evidence that AI memory mostly doesn't work.
GitOfThoughts treats an AI agent's reasoning tree as a git repository: each scored thought becomes a commit, scores become notes, and outcomes get tagged. Retrieval is just git log. The system makes an agent's reasoning replayable, diffable, and mergeable with other agents' histories at near-zero engineering cost. The researchers then ran a five-way comparison of memory substrates (none, markdown, vector, graph, and git) across two benchmarks and two model scales, with pre-registered replications.
The finding quietly undercuts the pitch behind AI memory systems: memory only helps when the retrieved case is nearly identical to the current problem (similarity above roughly 0.8), at which point the agent is copying an answer, not transferring a method. A model 4.5 times larger doubles that near-duplicate payoff but still cannot extract a generalizable approach from a worked example. The only lever that reliably moves accuracy across the board is test-time sampling.
That's a useful result for anyone building auditable AI systems, and a quiet rebuke to the vector-database companies selling memory as a path to smarter agents.