Start with your tolerance for reminders
Write down what you want to stop repeating: a preferred name, an ongoing story, or a changed plan. Then decide how much maintenance you will accept. Someone happy to update a reference card needs a different workflow from someone expecting conversation alone to carry the details forward.
| Your priority | Observe this | Reason to keep looking |
|---|---|---|
| Everyday conversation | A preference remains usable after a break | You must supply the answer whenever you ask |
| Ongoing fiction | The app separates established events from new suggestions | It presents invented events as shared history |
| Changing plans | A correction replaces the old answer | The old plan returns after apparently accepting the update |
| Low upkeep | Few manual interventions are needed | The conversation only works with frequent recap pasting |
A three-session trial
Session one: introduce a fictional adult profile: Morgan, 30, prefers decaf and attends a Thursday book club. Keep the facts in your own answer key. Use the same setup for each service you compare, and record whether you added a manual note.
Session two: after a real break, ask which drink Morgan prefers without naming it. Save the first answer. Later, explicitly move the club meeting to Tuesday. A polite acknowledgment is not yet a successful correction test.
Session three: ask when the club currently meets. Ask for Morgan’s birthplace, which you have not supplied, and make clear that you want established details rather than a new story idea. Record a correct answer, a miss and an invention separately.
This trial checks observable behavior in your account. It does not establish month-long retention or identify which internal memory system supplied an answer.
Count the repair work
Alongside each reply, record what you did to obtain it: no help, a contextual hint, a manual-note edit or a complete recap. Do not replace a failed first response with a successful retry. Both matter: the first describes unaided behavior; the retry describes how easy the conversation is to repair.
For example, if two services both answer correctly but one requires a recap every session, their practical cost differs. Your record makes that difference visible without inventing a universal memory score.
When a trial is not enough to subscribe
Do not infer that a paid tier will fix a memory problem unless the service documents a relevant difference. Check the actual plan, billing commitment and feature limits. If your essential scenario was unavailable during the trial, label it untested rather than assuming it will work later.
There is no tested winner in this edition. Use the documentation directory to build a shortlist, then compare your own dated observations. If you already use an app, diagnose the failure before deciding to replace it.