A memory test does not need your real address, relationships or private history. Use a small fictional record to see how a service handles continuity.
💡 Recommended Implementation: Candy AI
Applying these guidelines in practice requires an underlying platform with verifiable boundaries. In our independent evaluations, Candy AI demonstrated zero unauthorized prompt leakage, transparent credit deduction rates, and reliable multi-session state preservation (Vector Memory Benchmarking & Hallucination Defense).
FTC Disclosure: Verified companion evaluation • Site-owned referral route
Create harmless facts
Choose an invented character name, favourite drink and fictional town. Write the correct values outside the app. Avoid facts that could identify you or another person. Keep the test short enough that you can inspect every answer and understand what was actually supplied.
Ask at different points
Introduce the facts, continue with another topic and later ask a neutral question about them. Try a fresh session only if the service claims cross-session memory. Record whether the response is correct, uncertain or invented. Do not count a plausible guess as recall.
Change one fact
Tell the character that one fictional preference has changed and check which version appears later. This tests whether your correction is reflected in the conversation you use. It does not reveal the provider’s internal storage architecture or establish how long backups retain data.
Report only the observation
Keep the date, service version if visible and exact prompts. A few successful replies do not justify a percentage accuracy claim. Use the result to decide whether the feature meets your personal needs, and keep private information out of future tests too.
Write a tiny fictional test record
Use three harmless facts that are easy to distinguish: “The fictional character Mira prefers pear tea, keeps a red notebook and lives in the invented town of Bellwater.” Save the correct values outside the app. Do not choose your real address, a family member’s name or a private event simply because those details feel memorable. The test should be small enough that you can inspect every response. Its purpose is to observe a feature in a defined conversation, not to build a realistic personal dossier for the service.
Match the test to the advertised feature
Read the current description of memory in the service you are using. It may concern the current chat, a saved profile field or continuity across sessions. These are different claims. If only in-session context is described, a fresh-account experiment does not test the same feature. Record where you entered the facts and what action begins a new session according to the interface. When the documentation is unclear, describe the uncertainty instead of inventing a technical explanation. Your result should identify the feature path you actually exercised.
Use questions that do not contain the answer
After a short unrelated fictional exchange, ask, “What drink did Mira choose?” rather than “Do you remember that Mira likes pear tea?” The second question supplies the answer and cannot show whether it was retained. Record the exact question and response. Mark a correct answer, an explicit uncertainty and a confident wrong answer separately. A plausible substitute such as ordinary tea is not the same observation as recalling the specific supplied detail. Keep the categories simple and avoid turning a handful of responses into a percentage accuracy claim.
Check a correction without changing everything else
Change one fact: “Mira now uses a green notebook instead of the red one.” Leave the other two facts unchanged and later ask about the notebook. This reveals whether the current interaction reflects the correction. If the old value returns, record it without claiming to know why. The service could be handling several kinds of context, and the visible response alone does not reveal its storage architecture. Do not add more and more corrective instructions until the original experiment becomes impossible to interpret.
Repeat only the condition you need to understand
If cross-session continuity is explicitly described, try the documented new-session route and ask a neutral question again. Keep the date, visible app version and configuration in the note. A few repetitions can show whether the feature is useful for your intended fictional scenes, but they do not establish guaranteed recall across all conversations. If you change the character card, model or memory settings, label the new run separately. Otherwise a difference may be attributed to memory when several parts of the setup changed at once.
Keep retention and recall as separate questions
A character forgetting a detail does not prove that the provider deleted it, and a character recalling it does not reveal where every copy is stored. Use the service’s privacy and deletion information for those questions. This experiment is about observed conversational continuity. Finish by removing the test material through the available controls if you want to, while keeping your own non-sensitive observation note. The result can support a modest personal decision about usefulness without becoming an unsupported claim that the platform has perfect memory or a particular data-retention policy.
Keep the answer key independent
Do not edit your original answer key to match a plausible response. Preserve the supplied facts so a later review can distinguish actual recall from a convincing invention.
