Data
Conversational data is not retained.
Midium doesn't retain conversational data, ensuring maximum security and privacy. Prompts and generated responses are discarded. We do not train on them.
Cloud inference
Token-billed and dedicated cloud inference are processed to return a response, then discarded. We do not keep customer prompts or generated outputs as a product feature, log, or training corpus.
Midium Desktop
Midium Desktop runs inference and your local Vaults on your machine. Prompts, outputs, files and any data ingested or generated are not discoverable by others, including the Midium team. Hardware running Midium Desktop can operate with no network connection once models are downloaded, making truly air-gapped AI fully attainable.
Vaults persist until you delete them
Vaults are the exception. They are persistent context stores you write into. Vaults are encrypted end-to-end. Hosted Vaults stay until you delete the vault or its items. We do not use vault contents to train models. Local Vaults never leave the licensed machine.
What we do keep
We may process limited account, billing, and operational metadata required to run the service, for example token counts and member counts for invoicing. Contact form inquiries are stored in Midium's CRM so we can reply.
SOC 2 Type II targeted by January 15, 2027.
