Memory & control
AI companion memory vs. chat history: what is the difference?
Chat history is a record of conversations; saved memory is selected information used to personalise later responses. Keeping a transcript does not mean an AI can recall every part of it, and deleting one memory does not necessarily delete the underlying conversation.
Separate three layers of information
Implementations differ between products. Some summarise older material or retrieve selected records; others rely mainly on the current conversation. This distinction is a way to ask better questions, not a claim about the internal architecture of every app.
| Layer | What it means | What not to assume |
|---|---|---|
| Chat history | Messages or transcripts retained by the service. | That every old message is available to the model on every turn. |
| Current context | The information supplied to the model for a particular response. | That it contains your entire account history. |
| Saved memory | Selected details kept for use across conversations. | That the details are correct forever, or that every reply uses all of them. |
What is worth remembering?
Start with a harmless preference that makes future conversations easier: you prefer brief replies, you are practising a particular language, or you want to return to a personal project. Avoid treating memory as a secure place for passwords, financial details or information about people who have not agreed to share it.
Make the detail precise and revisable. “I am learning Japanese” says less than “For this month, I want to practise introducing myself in Japanese.” The second statement makes it clearer when the memory may need updating. Neither is an instruction the model is guaranteed to follow every time.
Check the controls, not just the reply
After asking an app to remember a test preference, open its memory controls. Does a corresponding record appear? Can you read it in plain language, change it and remove it? If the model says it remembers something but there is no visible control, ask the provider how that information is stored.
In Fari, saved memories can be reviewed, edited or removed in the app. Use those controls to correct a mistaken detail. A follow-up response alone is not a reliable audit: the model might still see the detail in the current conversation, or might guess correctly without retrieving a saved memory.
- Save one non-sensitive test preference.
- Inspect the saved wording and correct it if necessary.
- Remove the test memory through the app’s controls.
- Check the privacy policy separately for transcript and account deletion.
Memory deletion is not the same as account deletion
Ask which records a deletion control affects: a saved memory, a conversation, a journal entry or the whole account. These may have different retention rules. Removing a preference should not be presented as proof that every related record, backup or provider log has disappeared.
Fari’s privacy policy describes stored messages, journals, voice-call transcripts and memories. It also explains account deletion and the associated retention periods. Read that policy before sharing personal information; the short phrase “remembers you” is not a substitute for a data-handling explanation.
Treat continuity as a convenience, not proof of a relationship
Remembered context can make a conversation feel less repetitive. It can also be incomplete or wrong. Correct the app directly when it makes an assumption and keep important plans or decisions somewhere you can verify them yourself.
An AI can refer to a detail from yesterday without having human awareness or feelings. You should remain in control of what you share, how long you spend and whether you return. Choose the tool for the experience it provides, not for a promise that it understands everything about you.
Explore conversations and memory controls in Fari
Available on Android and in beta on iOS. Check the store listing and in-app terms for current access, prices and limits.
Explore Fari’s features
