How defensible is AI memory?

If I use ChatGPT and it updates its memory, then I throw the raw chat history into Claude -- how hard is it for Claude to reach "memory parity"?

(Assume there's easy import/export of chat history)

Also I’d assume data export will be legally mandated.

Working on Farcaster and Warpcast. Longer thoughts at https://dwr.email

If building from your own open-data product, the value primarily comes from secondhand modification of raw data, making it about as defensible as any strong analytics play.

GPT’s current advantage feels temporary, even if it’s very meaningful right now. My ideal scenario would probably be something like them acquiring a Farcaster data play rather than competing directly on custom models—a route that will likely never be cost-effective at scale.

OpenAI can explore 100 model iterations for every (one) that I could, leveraging extensive resources and experience. Meanwhile, Llama or similar projects can’t risk violating privacy again by training on closed data—so open-data sources have unique leverage here.

Also worth noting: the emotional and individual actor layer of AI remains nearly untouched— “I noticed you’re feeling [state] today. Consider speaking with [identity].”