Skip to content
The ReturnA Meliorly publication

AI & privacy

What should an AI habit coach know about you?

Useful context is not the same as unlimited context. A trustworthy coach should explain what it sees, why it needs it, how long it remains, and what you can remove.

One botanical branch passes through four focused vellum windows while unrelated branches remain outside in separate paper compartments.
Original editorial artwork for The Return.

Ask about four different layers.

Now

The current message

What you are asking the coach to answer right now.

Thread

Recent conversation

A bounded window that keeps the reply coherent without replaying everything.

Plan

Habit context

The cue, ordinary version, tiny version and recent pattern relevant to this habit.

Memory

Saved facts

Information retained for later use, which should be visible and removable.

Seven questions worth asking any AI coach.

  1. What enters a model request?Look for concrete fields, not “we use your data to personalize.”
  2. Can context cross between habits?A sleep conversation should not silently appear in a screen-use reply.
  3. Is chat history different from memory?Keeping a transcript does not mean every old message must be resent later.
  4. Can I inspect and delete memory?A control is meaningful only if you can find and use it.
  5. Does the model receive my identity?An email address or account identifier is rarely necessary to answer a habit question.
  6. Where does processing happen?On-device and cloud systems have different boundaries; neither label replaces a full explanation.
  7. When is AI the wrong tool?Everyday coaching needs a clear boundary around medical, crisis and specialist situations.

What a context receipt can look like.

Example request“Evening reading” Coach thread
IncludedCurrent message, recent same-thread turns, approved plan, confirmed relevant memory
ExcludedEmail address, account ID, other habit threads, rejected or deleted memory, unrelated free text
This is a product explanation, not a live display of anyone’s data.

Compare designs, not privacy adjectives.

Some systems draw on broad health and activity context. Google, for example, describes a Health Coach that can use connected health, fitness and conversation information. Apple documents a different path in which models can run on the device. Those are architectural choices, not simple good-or-bad labels.

Meliorly’s current boundary.

A model-generated Meliorly reply can receive the current message, a bounded recent window from the same habit’s Coach thread, the approved plan, limited momentum information and a small set of relevant memory notes you confirmed. It does not need your email address or another habit’s conversation.

Read the exact fields and controls on AI coaching and your privacy, then use the full privacy notice for storage, retention, export and deletion details.

Help us understand what is useful?

With your permission, Google Analytics measures visits and actions such as making a plan or downloading a workbook. It uses cookies. Your habit answers are never included. The site works fully with analytics off.

Read about website analytics