Inference APIs now return provider-sealed state only the issuing vendor can read
Earendil Engineering catalogs how inference APIs stopped returning portable transcripts: encrypted reasoning blobs, hosted searches whose source text never reaches the client, and conversation state keyed to provider-held IDs. OpenAI's Responses API retains those objects at least 30 days by default, Gemini's Interactions API 55 days on paid tiers and one day on free, and Anthropic's docs tie thinking blocks to the model that produced them. Portability now takes explicit work — store: false, a canonical local event log, client-side search tools — since a response ID is a foreign key into a database the application does not control.
Source: earendil.com ↗
A response ID is not a transcript (as the data is stored on the server), a ciphertext is not user-controlled stated (as the user cannot decrypt it), a list of citations is not the evidence that was placed in the model's context by a search result (as you cannot typically fetch th
Why this matters
- → AI conversation history no longer portable between providers — locked into vendor infrastructure
- → Users cannot fully export or audit what models saw, limiting control and reproducibility
- → Reasoning, searches, and state encrypted or server-stored, shifting session ownership to providers