How AI girlfriend memory works: scenes, preferences and corrections

What does an AI companion actually remember?

A reply can draw on recent messages, the character's identity, your profile preferences and relevant memory. These have different jobs: identity establishes who the companion is, preferences describe how you like to interact, and memory helps an ongoing relationship carry forward.

A summary compresses information. An old detail may be omitted, or the model may interpret it incorrectly. Treat memory as a useful continuity aid and check the details that matter instead of assuming every line of chat will be remembered forever.

Should you use shared or private scene memory?

Use shared memory when several scenes belong to the same relationship. Use a private scene when you want its memory kept within that scene. This distinction concerns memory sharing between your scenes; it does not make a shared scene public or share it with other users.

Keep a new scene's setting explicit. A remembered visit to a café should not move a conversation away from the beach you are currently describing. Give the current place and activity in the opening message.

Today we're walking by the sea. Keep our preference for quiet places, but this is a new scene, not the café from yesterday.

A five-minute memory audit before you trust a detail

Use a low-stakes fictional preference rather than private or identifying information. Ask the companion to acknowledge the preference, continue the scene for a few turns, then start a fresh message and refer to it naturally. This tests whether the fact is available in the current context without turning a personal secret into a diagnostic.

If the answer is wrong, check the saved memory before changing the character. A missing fact, an incorrectly summarized fact, and a deliberate scene boundary are different problems. Record which one you saw so that one bad reply does not lead you to over-edit the whole relationship.

1. Choose one harmless preference, such as a fictional favorite drink.
2. State it once in plain language and ask the companion to confirm it.
3. Continue the conversation without repeating the fact every turn.
4. Refer to it later in a new scene and compare the response.
5. Correct or remove the saved memory only after checking what was stored.

I like jasmine tea in our fictional café scenes. Please remember that preference, but do not treat it as a rule for every location.

How do you fix a mistaken memory?

Correct the fact directly, then review the saved memory in Settings. A clear correction is more useful than adding a competing fictional explanation. User-confirmed facts help keep the intended version explicit.

1. Identify the incorrect detail, such as a name or preference.
2. State the replacement in one clear sentence.
3. Review and correct or delete the saved memory in Settings.
4. Continue the conversation and check whether the next reply respects it.

What memory cannot promise

Memory is a continuity aid, not a verbatim archive and not a guarantee that every generated reply will agree with every earlier detail. A new scene, a long conversation, ambiguous wording, or a competing instruction can change which context is most visible to the model.

For sensitive information, use the smallest useful detail or leave it out. Product memory controls help you inspect and correct saved facts, but they do not turn an AI conversation into a secure personal records system. Check the privacy policy before sharing information you would not want retained.

What is a realistic continuity test?

Use a fictional, low-stakes detail, such as a favorite drink. Establish it naturally, continue the scene, and revisit the topic later. Check relationship tone as well as factual recall. One correct answer does not prove permanent memory, and one mistake is a reason to inspect the saved fact rather than restart the whole character.