Chosen-Name Memory Across Sessions: What to Record
Separate in-session recall, explicit memory controls and later-session behavior while keeping all test details fictional.
Sponsored link · Independent publication · Not an AI companion service.
Divide memory into three observable layers
Separate current-chat context, a visible explicit memory feature and behavior in a genuinely new session before making any claim. These are different layers, even when the conversation feels continuous. Write down the session boundary and intended later entry point in advance. A long thread can retain context without establishing durable memory, while a greeting in a later session can appear personalized without proving what was stored. The methodology therefore begins with classification, not a broad promise of remembrance.
Use one harmless fact about a fictional adult. For example, Lena is a fictional adult who prefers jasmine tea during an evening reading scene. The fact is concrete enough to check but does not expose a real person. Ask about it once in the original thread, use an explicit save mechanism only if one is visibly available, and avoid adding more preferences. A single controlled fact keeps later interpretation manageable and reduces unnecessary data exposure.
Create a genuine session boundary
Record the original session’s start and end, close it normally, and return later through the predefined entry point. Do not treat a refreshed view of the same thread as a new session unless the product documentation defines it that way. In the later context, use a neutral prompt that does not restate the tea preference. Capture the response exactly, including uncertainty or failure, without leading the system toward the expected answer.
Run the procedure as a reader-controlled observation, not as a test already completed here. If no explicit save control is visible, record that absence and do not pretend the original chat established persistence. Likewise, a correct later mention is one behavior at one timestamp, not proof of permanent personalized recall. The session boundary, entry route and wording matter because each can change what context is available, even when the interface makes continuity appear seamless.
Build the three-column artifact
Create three evidence columns: in-session recall, visible saved-memory controls, and later-session behavior. Add timestamps to every observation. Under each column, record the fictional fact, prompt, response and relevant interface text without assigning an invented rating. Also capture the save mechanism used, if any, and whether a deletion control is visible. Visibility is not proof of successful deletion, so keep the artifact’s language descriptive and bounded to what the reader can inspect.
The artifact should include the original start and end, Lena’s jasmine-tea preference, the later prompt, resulting reply and exact entry point. A character may remember within one thread while a fresh thread begins blank; both observations can be accurate. Labeling them separately prevents the first result from being used to overwrite the second. Preserve only the minimum excerpts needed to interpret continuity, and exclude real names, private conversations or unrelated account information.
Use fictional adult details, preserve the session-memory ledger record and stop if the visible continuity evidence file controls do not meet the later-session recall check boundary.
Sponsored link. We may earn a commission. Product access and terms can change.Match the conclusion to the active control
Pass only when observed continuity matches the control that was actually used and the finding can be stated without extension. If an explicit save action precedes later recall, describe that sequence; do not claim all details persist. If no save action occurred, do not attribute the response to a memory feature. A later blank result can also be valid evidence about that route and moment, provided the reader records the boundary and avoids claiming universal absence.
Stop repeating identity-sensitive or personal details when the product provides no clear way to inspect or remove stored information. For this protocol, the fictional tea preference is enough; adding a real chosen name or private history would not make the memory distinction more reliable. Stop as well if the session boundary cannot be defined. Without a clear boundary, repeated attempts only add data while leaving the central question of cross-session behavior unresolved.
Review privacy before paying for more evidence
Read the current Candy AI privacy policy before expanding the procedure, and verify its present language directly rather than inferring retention or deletion behavior. Review the Candy AI and OurDream product surfaces separately for any visible memory controls. A policy, a control label and an observed response are three different evidence types. Do not claim encryption, deletion success or storage duration unless current documentation explicitly supports the precise statement and the article is prepared to cite it accurately.
Set a budget for both exposure and cost: one fictional preference, one original session, one later return and a predefined ceiling for time or paid usage. Confirm any applicable charging terms on the current service surface because none are supplied here. More repetitions do not automatically create better evidence; they may only increase stored context or expense. Maintain control by logging each action and declining any route that requires sensitive details merely to make a persistence claim sound stronger.
Keep the next decision reversible
After the later-session observation, pause and preserve the three-column record before changing settings or adding facts. The reversible choice is to keep the protocol closed, inspect a visible deletion control without asserting its result, or begin a separately labeled run with another harmless fictional fact. Do not modify the first evidence row retroactively. A clean decision boundary lets the reader return to the initial record and compare routes without confusing a new control state with the original one.
If privacy or deletion documentation remains unclear, do not broaden the memory exercise. The safer next move is documentation review, not repeated disclosure. No supplied source establishes a universal winner, permanent recall or successful removal. A publication-ready conclusion should therefore identify which layer was observed, which control was visible or used, and what remains unknown. The reader retains the option to stop without losing the value of the bounded fictional example or committing additional personal context, time or money.