You stay in control: sensitive identifiers are masked before transmission
Using ChatGPT, Claude, or Gemini directly means sending whatever you type as is: a pasted IBAN in a question, a code received by SMS, a photographed ID. Happy Assistant steps in: sensitive data is detected and masked before calling the model, your memory stays in the app, and nothing is used to train AI.
Key takeaways
- Card numbers, IBANs, Social Security numbers, verification codes, passwords, and API keys are masked before being sent to the model.
- If your message visibly contains sensitive data, the app warns you and offers to send it masked.
- Your memory (family, health, work, preferences) is stored in the app, viewable, and erasable — not in the model.
- The sent content is not used to train the underlying models.
Phrases that work
- What do you know about me?
- Forget what I just told you about my work.
- Delete everything you've remembered about my health.
- Don't remember anything from this conversation.
What is filtered before being sent to the model
Before a message leaves the app, its text is analyzed and sensitive patterns are replaced by a marker. The model receives the sentence and its meaning, not the secret.
- Credit card numbers, validated by the Luhn check to prevent false positives.
- IBANs and account numbers.
- Social Security numbers (France, US).
- CVV cryptograms and card expiration dates.
- Verification codes and one-time passwords (OTPs) received by SMS or email.
- Plain text passwords.
- API keys and authentication tokens.
Filtering applies to the current message, the conversation history sent back to the model, the content of attached text files, the analysis of your emails, and memory extraction. A detected secret is never saved as a permanent memory.
A warning before sending, not after
- You write or dictate a message containing an IBAN, a code, or a card number.
- The app detects it on-device, before any network call.
- An alert indicates what type of data was recognized.
- You choose: send a masked version, correct your message, or cancel.
Attachments follow the same prevention logic: a photo or PDF triggers an explicit reminder before sending, because a scanned document (bank statement, prescription, ID) cannot be reliably masked automatically.
Happy Assistant compared to using an LLM directly
| Happy Assistant | Direct ChatGPT / Claude / Gemini |
|---|
| Pre-send filtering | Automatic: cards, IBANs, codes, passwords, keys | No filtering added by Happy Assistant |
| Memory | Stored in the app, categorized by domain | Depends on the provider, plan and account settings |
| Deletion | Memory by memory, with one gesture | Depends on the provider, plan and account settings |
| Training on your data | No | Depends on the plan and account settings |
| Attachments | Systematic warning before sending | Depends on the interface used |
Your memory lives in the app, not in the model
Useful facts — your children's names, allergies, work city, fitness goals — are saved in your space, organized by domain, and injected into the conversation only when needed to answer. The Memory screen displays them all; each can be edited or deleted individually.
- Domains marked as sensitive (health, finance) can be excluded from answers.
- No data is shared with other users or sold.
- Connected mailbox credentials are encrypted and never transmitted to the model.
- No server logging of your messages, photos, or documents content.
Details regarding processing, sub-processors, and your rights are available in the app's privacy policy.
What the feature does not do
- Masking applies to text: a photo or a PDF cannot be reliably redacted, hence the warning displayed for every attachment.
- No filter is perfect: an unusually formatted identifier might slip through, so the pre-send warning remains your safeguard.
- Processing goes through third-party AI providers: content passes through their servers for the time of the response, without being permanently stored or used for training.
Frequently asked questions
Are my messages used to train AI?
No. Calls go through professional APIs that do not use the transmitted content to train their models.
What happens if I paste my IBAN by mistake?
The app detects it before sending, warns you, and offers to send a masked version. Even if you confirm sending, server-side masking applies before calling the model.
Can I delete what the assistant has remembered?
Yes, memory by memory from the Memory screen, or by asking it verbally. Deletion is immediate and permanent.
What if I take a picture of a sensitive document?
An image cannot be reliably redacted: the app warns you before sending, and it is best to avoid taking pictures of a bank statement, prescription, or ID.
Where is my data hosted?
In Happy Assistant's application infrastructure, with user-restricted access. Hosting details and retention periods are in the privacy policy.