We originally assumed OpenAI's data export was just a flat array of messages. It’s not. It’s actually a node tree (because whenever you edit a prompt or click "regenerate response", it branches off).
When we naively parsed this into linear commits, our timelines looked completely insane. The agent would make a great architectural decision, then suddenly time-travel back 5 minutes to suggest a terrible alternative, and somehow "merge" both into reality. It looked like a Git rebase gone horribly wrong.
We had to build a DAG (Directed Acyclic Graph) resolver and a deduplication layer just to flatten OpenAI's multi-verse of regenerated responses into a single, sane Git history before committing it to the repo.
Turns out, AI memory isn't just lossy it's non linear. Anyone else here suffered through parsing OpenAI's export JSONs?