A privacy mode sounds simple, but different AI companion apps can mean very different things by “incognito.” Users should check exactly which data flows change.

Incognito should define memory behavior

A private session should clearly state whether conversation details can become long-term memories. If the mode only hides a chat from the visible history while still updating personalization, users may misunderstand its purpose. History and retention are separate.

A conversation can disappear from the interface while remaining temporarily on servers for safety or operations. Apps should explain retention rather than using vague “private” labels.

Training controls matter too

Users should know whether private sessions are excluded from model improvement or analytics. This is separate from whether the conversation is remembered by the companion.

Media needs its own rules

Uploaded photos, voice notes and generated images may have different storage lifecycles from text. Check whether media is deleted with the private chat. Test the mode.

Share a harmless temporary preference in a private session, start a normal session later and see whether the companion recalls it. Do not use sensitive information as a test. For broader settings, see AI Companion Privacy Settings.

Understand what is stored versus what is inferred

Some products retain explicit user inputs, while others also create summaries, embeddings or preference profiles. User-facing controls should make the important consequences understandable even if the implementation details stay hidden. A practical question is simple: after deleting or correcting a detail, does future personalization stop using it?

Review third-party and creator access separately. For creator-backed AI, users should know whether the human creator or management team can see private conversations. That access model should be explicit rather than implied by the creator’s public identity.

Check privacy where the data is actually exposed

Privacy is not only a policy-page question. Test lock-screen notifications, shared-device behavior, account switching, export files and deletion controls. A product can have a reasonable privacy statement while still exposing sensitive context through previews or leaving old data available on another device. Users should be able to understand which permissions are required for the core experience and which are optional enhancements.

Verify deletion end to end. Deleting a conversation, a memory, an uploaded asset and an account may be different operations. The interface should explain the scope of each action and prevent removed information from reappearing through summaries, caches or synchronization.

For sensitive media such as faces or voices, it is especially important to know whether the source file, derived model and generated history are managed separately.

Use the smallest permission that still delivers the feature

Location, contacts, photos and microphone access can each improve an AI experience, but they should not become default requirements when a narrower permission would work. Review whether the app offers limited photo selection, approximate location or one-time access where appropriate. The product should also keep working in a reasonable fallback mode when an optional permission is denied.

Check synchronization and revocation. If privacy settings are changed on one device, verify that the new choice reaches other signed-in devices. A revoked permission or deleted memory should not remain active because another client is using stale state.

A final long-term check

Repeat the same test after several days and after changing one setting. AI products accumulate memory, synchronization state and billing or permission history, so problems may appear only after the first-session novelty has passed. The strongest products remain understandable as history grows: users can see what changed, correct the relevant state and continue without rebuilding the entire relationship.

Bottom line. A useful privacy mode should explain memory, visible history, retention, training and media separately so users know what “private” actually changes.