Start with a brief you would want reused
A first name can make a greeting familiar, but the helpful context is often more practical: the outcome you are working toward, the constraints you keep repeating and the way you like options presented. Share one detail that changes the next result.
For example, “I prefer afternoon appointments and two concise options” is actionable. It does not require an assistant to collect your entire schedule or infer a permanent lifestyle from one conversation. Keep the first brief short enough that you can recognize an error.
Distinguish facts, preferences, exceptions and guesses
An assistant may combine information you supplied with a working assumption. Those should not look the same. A temporary exception should not quietly replace a usual preference, and an inference should not be presented as something you explicitly told it.
| Kind of context | Example and useful treatment |
|---|---|
| A supplied fact | The appointment is on Thursday; use it for this task. |
| A preference | Usually avoid early starts; apply when relevant. |
| An exception | This trip needs an early departure; keep the exception local. |
| An assumption | Travel may take forty minutes; label and verify it. |
Try a harmless correction
Use invented details to evaluate memory controls before sharing something sensitive. Tell the assistant a simple preference, ask it to summarize what it will use and then correct one part. See whether the next draft reflects the correction.
A successful correction should reduce repeated explanation. If the product has separate saved-memory and conversation-history controls, check which one you changed. Ask what remains in uploaded files or connected services; the same word, “delete,” can describe different scopes.
Look for a change in the work, not just the greeting
After an assistant learns that you prefer fewer options, a useful next comparison should be shorter. If you have a fixed budget, a plan should respect it or explain why it cannot. If you correct a constraint, the proposed result should change coherently.
Personalization should also know when a detail is irrelevant. A travel preference need not influence an unrelated writing task. Ask the assistant to explain which saved details affected the result when that would help you judge its reasoning.
Keep a way to inspect, revise and remove
Before relying on long-term memory, look for a clear explanation of where preferences are kept, who can access them and how you can stop using them. Ask whether removing a connection and deleting saved context are separate operations.
Useful anonymous behavior matters too. You should be able to explore a product or complete a small task without pretending to be a returning customer. A service should not claim to recognize you from a name or a page visit you did not choose to associate with an account.
Twin’s personalization direction
Twin’s product direction is a one-to-one helper shaped by the context, goals and permissions a person chooses to share. Correction, removal and keeping each person’s information separate are requirements of that direction.
Those requirements do not establish that every future memory control is available in the current product. The customer operator and optional app are in development. The website’s current privacy policy explains the public website experience; an early-access introduction is not consent to collect your personal history.