Zero retention and local AI: trace the actual data path
A retention promise needs a scope
Ollama’s FAQ distinguishes local operation and cloud-related features. A statement about model requests does not establish what a surrounding application, proxy or support system retains. “Local” describes where an operation runs; “zero retention” describes a particular handling commitment. Neither phrase alone documents the complete path of a user’s information.
Trace one synthetic request
| Stage | Record |
|---|---|
| Browser or client | Storage, extensions, account and request destination |
| Application | Prompt construction, retrieval sources and application logs |
| Model service | Exact endpoint, configuration and applicable retention terms |
| Monitoring | Error reporting, analytics, proxy logs and their destinations |
| Recovery | Backups, exports, support records and deletion process |
Use invented sample information with a distinctive test marker. Inspect available configuration and authorised logs for that marker. Do not send a real secret merely to test whether it leaks. An absent log entry is not proof of no retention where logging or access is incomplete.
A narrow finding is more useful than a slogan
Imagine a local model request that stays on one machine while its application sends error reports to a hosted service. The model path is local; the whole workflow is not proven local. Record what the error payload contains and what setting controls it. Conversely, a network connection for downloading a model is different from transmitting a prompt; classify the operation before drawing a conclusion.
Recheck after an operational change
Repeat the map when the model provider, cloud feature, application version or logging policy changes. Keep the dated terms and configuration with the test record. For sensitive work, require evidence for the unknown stages before expanding use. This is an audit method, not a completed inspection of a reader’s system or a blanket guarantee about a provider.
Sources and scope
- Ollama FAQ — Product-specific local and cloud behaviour; check the version and configuration actually used.
Sources checked October 10, 2026. AI-assisted research and drafting. The examples, calculations and worksheets are Luminesca editorial work; they are not measured trials, first-hand reporting or an independent certification. Removed claims that every local model or zero-retention option eliminates all copies. Report a specific correction with its source.