Imagine a fictional companion remembering that your story started in a closed record shop. It feels personal. Then she puts the next scene in a café. Did the memory fail, or did she just invent a new location? Understanding the difference helps you judge the app and repair the conversation.
This explainer uses public documentation checked on 3 October 2026. The checks below are suggestions, not results from a paid account.
A callback can come from the message you just sent
Compare two questions: 'Remember our evening in the record shop?' and 'Where did our story begin?' The first gives away the location. A reply about records can be convincing even if the app retrieved nothing. The second checks whether an earlier detail is available without your repeating it. The app may still supply that detail from chat history, so a correct answer does not establish separate saved memory.
The reply is generated from information made available to a language model. Its active context can contain character instructions, recent messages and other material supplied by the app. Anthropic's context documentation explains the general distinction between this working context and the data used to train a model. It does not identify which model any companion service uses.
That distinction matters when a reply sounds certain. Confidence in the wording does not tell you whether a fact was available, retrieved correctly or guessed. Keep the question and the answer together when judging a callback.
Saved history and usable memory are different
An app can keep a conversation without putting every old message into each new request. One documented engineering approach is to condense earlier turns into a summary; another is to retrieve relevant material when needed. Anthropic describes both approaches. These are possible designs, not evidence that every companion app uses the same one.
A useful summary might preserve 'the story began in a record shop' but lose the name of the album. More text is not automatically a cure: the context documentation also describes declining recall as context grows. For the reader, the practical question is whether the app carries the details your story needs, including a correction made later.
| Documentation checked | What it says | What it does not establish |
|---|---|---|
| OurDream public landing | A large context window carries history and details across days and weeks. | Complete recall of every turn, or a measured success rate. |
Those are provider statements checked against OurDream’s public page. They do not establish complete recall or a memory winner. If continuity matters to you, browse OurDream’s characters and run the same small check.
Give her something she can act on
Character instructions help shape the reply. In the general engineering documentation, clear instructions and examples give a model usable guidance. For a fictional companion, 'confident' leaves room for many interpretations. 'She proposes a plan, asks one question at a time and admits when she misreads a joke' describes observable behavior.
Change one instruction and read the next few replies. Does she follow it when the conversation moves on? If you need a lasting profile change, look for an actual editor. A sentence in chat may affect the current exchange without changing a stored character setting. Check the editor’s current fields before treating a conversational instruction as a saved profile change.
The photo is another result to check
Good conversation and consistent pictures solve different problems. Image tools can use text and visual references to guide a generation. Hugging Face's IP-Adapter documentation gives one example of reference-based techniques for faces and style. It also notes that preserving facial details is challenging. This does not establish which technique a partner uses.
For your own fictional character, compare the face in several outputs rather than judging one attractive image. If reference controls exist, keep the same approved reference and change the outfit or location separately. If there is no reference control, compare outputs from the same character without assuming photo uploads are supported. Save the request when a result drifts: otherwise it is hard to tell whether the change came from your instructions or the generator. Our AI waifu video guide separates saved workflows from features that still need an account check.
Voice needs its own check too. Text-to-speech systems generate audio from text, as shown in Coqui's synthesis documentation. An app's voice message and live call can still be different features. Listen for intelligibility, natural pauses and whether the response follows what you said; a list of voice options cannot answer those questions.
A small continuity check before paying
- Establish two harmless fictional details, such as a meeting place and an unfinished plan. Save the exchange.
- Return in a later session. Ask about one detail without putting the answer into the question.
- Correct one detail explicitly: 'The shop was closed when we met; please keep that version.' Ask about it again after the conversation moves on.
- Record the plan, character, date and replies. Treat this as a small sample, not a reliability score.
Treat this as a continuity check, not proof of a separate memory system. Reopening a chat may carry its earlier messages into the active context; a correct answer does not reveal how it was retrieved.
Skip the upgrade if it promises the feature you need but you cannot establish what the plan includes. A service that suits short chats may still disappoint someone who wants a long continuing story. Decide what must carry over before you test; otherwise an entertaining reply can distract you from the missing detail.
Finally, treat visibility, storage and model training as separate questions. Memory alone proves none of the privacy terms. Read the named service's policy and keep identifying details out of a test conversation.
Frequently asked
Why does my AI girlfriend forget something I told her?
The detail may be absent from the active context, omitted from a saved summary or misunderstood. Ask about it without supplying the answer, then correct it explicitly. If it disappears again in a later session, keep the examples for support; one failed reply cannot identify the internal cause.
Does long-term memory mean she remembers every message?
No such guarantee follows from the label. Providers can retain history while supplying only selected information to a reply. OurDream describes a large context window. That description is not an independent measurement of complete recall.
Can I change her personality by chatting?
You can give instructions and corrections in a conversation. Whether a change persists in the character profile or future sessions depends on the app's controls. For a lasting edit, use the creator or profile editor when one is available and check the result in a fresh exchange.
Why does her face change in generated photos?
Image generation has its own identity and scene-control problem. A text description or reference can guide an image without fixing every facial feature. Keep an approved fictional reference and change one scene detail at a time when the app supports those controls.
Does a private chat mean the provider never processes it?
Private visibility and provider processing are separate questions. Check the named service's policy for collection, service providers, training use and deletion. Do not treat a memory feature or a private badge as proof of encryption or a no-training promise.