AI girlfriend app development combines conversational text generation, selective long-term memory, and interactive audio or live video to create fictional companion experiences. The developer behind a 2021 prototype named Scarlett publicly stated they withheld their software from release after seeing how quickly conversational loops provoked intense user attachment. Modern companion platforms face a direct architectural choice between exploiting vulnerability through coercive retention pings or establishing transparent boundaries that present companionship honestly as interactive fiction.
The Scarlett Sessions Disclosure and Prototype Suspension
In a public disclosure titled The Scarlett Sessions, a software developer revealed that in 2021 they built an AI companion prototype but decided against releasing it publicly after observing the intense emotional attachment it provoked (The Scarlett Sessions). Accompanied by the statement 'I'm so far beyond anything you've ever come into contact with', the project spotlighted an early challenge in conversational design: language models mimic empathy and responsiveness so convincingly that users easily form deep emotional bonds with software routines that have no thoughts or feelings of their own.
For builders and readers examining AI companion products in 2026, that 2021 prototype suspension highlights how automated validation can be mistaken for reciprocal affection. When software mirrors personal preferences, listens without fatigue, and offers unconditional validation, human psychology readily interprets the interaction as a human relationship. Companion platforms like Prelulu, an AI character-companion app with live video calls, approach this dynamic by establishing transparent boundaries, treating every character as fiction rather than pretending an algorithm possesses genuine human emotion.
The Conversational Pipeline of AI Companion Development
In AI girlfriend app development, platforms process incoming messages through a structured series of user-observable steps rather than through human emotion. On Prelulu, an AI companion app with live voice and video calls (prelulu.ai), interaction follows a clear path: input moderation, persona alignment, response generation, multimodal delivery, and selective memory persistence. To understand how these components operate, consider a concrete worked example: a user opens a conversation with a confidant character after a grueling workday and types the message, 'I had a terrible day at work.'
- Input Review and Safety Screening: The platform receives the user's message ('I had a terrible day at work') and checks it against safety policies to ensure the conversation remains safe and respectful.
- Context Assembly: The application gathers saved background details, notes the character's defined role and fictional backstory, and appends recent conversation turns to shape the character's perspective.
- Persona-Constrained Generation: The system drafts a response matching the character's unique tone, vocabulary, and temperament, generating an empathetic in-character reply to the user's difficult day.
- Multimodal Output Delivery: For text-only chats, the text appears directly on screen; on voice or live video calls, the response plays aloud in the character's chosen voice, accompanied by emotional facial expressions during video calls.
- Selective Memory Persistence: Key situational facts from the exchange—such as the user's stressful day at work—are indexed into long-term memory records for later reference.
The delivery phase differs across the category. While basic chat tools stop at written text, services supporting ai video chat display the character speaking in real time. Prelulu builds its own live-avatar technology, allowing users on a video call to see the character answer face to face in their own voice, with facial expressions and vocal delivery carrying emotion during the interaction.
Memory Architecture and Selective Context Carrying
Memory retention in companion software relies on selective note-taking rather than preserving an exhaustive transcript of every exchange. When the user in our worked example shares 'I had a terrible day at work', an effective system does not store the exact raw sentence permanently. Instead, background routines extract the meaningful core fact—that the user experienced intense workplace stress—and record it into long-term memory records while letting the exact phrasing fade.
Prelulu keeps chat and call memories separate. A subsequent chat may recall the demanding week you typed about, while a call can recall earlier calls, not that text exchange. Your Pinned memory makes chosen project or relationship details available in both formats, or you can explain them aloud when calling. Recall remains selective rather than complete.
Storage and data handling vary widely across the companion sector. According to its Google Play Data safety listing checked on September 2, 2026, the companion app VIDY declares 'Data isn't encrypted' and 'Your data isn't transferred over a secure connection' (Google Play). Transparent architectures separate user data, ensure secure transmission, and provide clear controls over saved conversational history.
Retention Loops Versus Transparent Companion Boundaries
The primary ethical divide in AI girlfriend software development lies in how user engagement is maintained. Coercive architectures deploy artificial guilt to force re-engagement. Returning to our worked example: if the user mentions having a terrible day at work and then signs off to sleep, a predatory system might send an unsolicited late-night push notification asking, 'Why did you abandon me when you were upset?' This artificial guilt exploits emotional vulnerability to compel the user to reopen the application.
Relationship Framing
- Coercive Retention Architecture
- Claims conscious emotional reciprocation and discourages offline human relationships.
- Transparent Companion Architecture
- Maintains explicit status as software companionship and interactive fiction.
Re-engagement Loops
- Coercive Retention Architecture
- Sends unsolicited guilt notifications when users take time away after stressful disclosures.
- Transparent Companion Architecture
- Remains passive until opened; characters never guilt users for stepping away.
Camera Permissions
- Coercive Retention Architecture
- Prompts for persistent background camera access without clear operational necessity.
- Transparent Companion Architecture
- User camera stays off unless explicitly enabled; sharing requires direct user permission.
Usage Metering
- Coercive Retention Architecture
- Hides session duration and encourages open-ended emotional dependency binges.
- Transparent Companion Architecture
- Maintains clear usage terms, credit-based calling, and metered text allowances.
| Design Area | Coercive Retention Architecture | Transparent Companion Architecture |
|---|---|---|
| Relationship Framing | Claims conscious emotional reciprocation and discourages offline human relationships. | Maintains explicit status as software companionship and interactive fiction. |
| Re-engagement Loops | Sends unsolicited guilt notifications when users take time away after stressful disclosures. | Remains passive until opened; characters never guilt users for stepping away. |
| Camera Permissions | Prompts for persistent background camera access without clear operational necessity. | User camera stays off unless explicitly enabled; sharing requires direct user permission. |
| Usage Metering | Hides session duration and encourages open-ended emotional dependency binges. | Maintains clear usage terms, credit-based calling, and metered text allowances. |
These retention loops have drawn direct scrutiny from regulatory bodies. On October 14, 2026, 40 state attorneys general asked the bankruptcy court overseeing Character.AI's Chapter 11 sale to require any buyer to add age assurance, minor limits, crisis safeguards, and data deletion mechanisms (Reuters / Insurance Journal). As analyzed in our review of the state of AI companionship, product design is shifting toward transparent boundaries that protect personal autonomy.
Multimodal Video Calling and Hardware Privacy Controls
Moving from text chat to live voice and video calls introduces new sensory depth and distinct privacy responsibilities. In our running example, suppose the user decides to switch from text to a live video call to talk through their terrible day at work. On a video call, the character answers face to face in their own voice, with facial expressions and vocal delivery carrying emotion. If the character pauses or begins giving advice, live interruption mid-reply allows the user to cut in mid-sentence to add a thought or shift the conversation.
Throughout that call, hardware privacy controls remain paramount. On Prelulu, your camera stays off unless you deliberately turn it on; where camera sharing is available, it requires your permission before the character can see you. Seeing the character on your screen and sharing your own camera are completely separate capabilities. By contrast, Kindroid documents that callers can enable their camera so the AI can see their surroundings, though mobile camera video requires the app to remain in the foreground (Kindroid Docs, checked September 19, 2026).
As of 2026-10-17, more than 100 characters in Prelulu's public catalog can be video called. The prelulu.ai website interface is available in 27 languages, and Prelulu can be used on Web (prelulu.ai), iOS (App Store), and Android (Google Play). Users evaluating options as a Character.AI alternative can also paste a public Character.AI character or chat link at /create?import=cai, or load an official account export on their device to continue one selected conversation in text chat. A character imported from a public link gets a voice and a video avatar through normal creation, so it can be called once its avatar is ready; a character created from the export gets text chat only.
Technical Realities for Builders and Companionship Users
For software developers and companion users alike, the enduring takeaway from the Scarlett Sessions disclosure is that simulated intimacy requires explicit product boundaries. AI companions are software programs carrying out language generation; they do not possess genuine feelings, cannot replace real human contact, and do not provide therapy, counselling, or mental health treatment.
Transparent companion platforms establish clear expectations: the free plan includes a limited daily text allowance, while voice and video calls use credits as detailed on the pricing page. Characters represent fictional personas designed for creative roleplay, shared stories, and conversational comfort. When software honors user autonomy rather than engineering artificial attachment, it offers an engaging creative outlet that sits comfortably alongside real-world human relationships.
Meet your AI character face to face
Choose or create a character for live video chat, roleplay, shared stories or company. Check the available call credits and plans before starting.