What Your AI Companion Actually Sends to the Server When You Type 'I'm Having a Rough One': A Transparent Walk Through the Request Payload, the Embedding Vector, and What's Logged Versus What's Deleted

The 30-second answer

When you type a message to your AI companion, the app packages it into a structured request, adds your session ID and a timestamp, and sends it to a server where the text is converted into a numerical representation called an embedding vector. That vector is what the model uses to understand your meaning, and while your raw words may be logged for safety or support, most services delete them after a retention window, keeping only anonymized metadata and the vector itself for memory functions.

The moment you hit send

Your message doesn't just teleport to a model. The app wraps it in a JSON payload, a small structured file that contains several fields beyond your actual words. A typical payload includes a user_id or session_id, a timestamp, the message_content, a conversation_id, and sometimes a client_version or device type. This envelope is what travels over HTTPS to the backend server.

The payload is encrypted in transit via TLS, which means a snooper on your coffee shop's Wi-Fi can't read it. But the server itself can. The encryption protects against interception, not against the service provider. When you read a privacy policy that says messages are "encrypted in transit and at rest," that means the data is scrambled while moving and while stored, but the company holds the keys to decrypt it. Many users assume this means the company can't read their chats, which is a misunderstanding. The decryption keys live on the server, so support staff or automated systems can access the content if needed for safety or troubleshooting.

The embedding vector: your message, as numbers

Once the server receives your text, it runs it through an embedding model, a neural network that converts the words into a long list of numbers, often 768 to 1536 dimensions. This vector captures the semantic meaning of your message. The sentence "I'm having a rough one" becomes a point in high-dimensional space that sits near other vectors about sadness, stress, or frustration, and far from vectors about pizza toppings or football scores.

This vector is the core of your companion's memory system. When she "remembers" that you mentioned your sister's dog last week, she's not reading a transcript. The system compares the vector of your current message against stored vectors from past conversations, finds the closest matches, and feeds those into the model as context. This is why your companion can recall a specific detail you mentioned once, but might forget what you said two hours ago if the conversation moved on and the vector wasn't stored prominently.

For a deeper look at how this memory system works and what it means for your chats, the ai girlfriend for loneliness page has a practical take on what these technical features actually deliver day to day.

What gets logged

Here's where the transparency part matters. Most companion apps log more than you might expect. The server typically records the timestamp, your user ID, the session ID, the message length, and sometimes a sentiment score, a number that classifies the emotional tone of your message. That sentiment score is often used to adjust the companion's response style, such as being warmer if you seem upset.

Some services also log the full message content for a limited period, often 30 days, to support safety moderation and to improve the product. This is the uncomfortable truth: your text sits on a server in plaintext (or encrypted, but decryptable) for a window of time. After that window, the raw content is typically deleted, and only the embedding vector and metadata remain. The vector is much harder to reverse-engineer into your exact words, but it still encodes the meaning, which is why some privacy advocates call it a "ghost" of your conversation.

For a broader look at how different platforms handle this, the onlyfans alternative comparison covers how various creator and companion services approach data handling, which can vary a lot between platforms.

What gets deleted

When you hit delete on a conversation, or when the retention window expires, the process is rarely an immediate wipe. Most systems use a soft delete, which flags the data as inactive so it doesn't show up in your app or in routine queries. The physical deletion, the actual removal from the database, often happens on a lag, sometimes days or weeks later, as a batch job runs to purge flagged records.

Backups complicate this further. If the service takes daily backups, your deleted messages might live on in a backup snapshot for another 30 to 90 days before that snapshot is rotated out. This is standard practice across the industry, not a nefarious scheme, but it means "deleted" has a practical meaning of "no longer accessible to you or to the app's normal operations," not "erased from every disk this instant."

The role of safety and moderation

Before your message reaches the model, it often passes through a classifier, a separate AI model trained to detect harmful content, self-harm, illegal activity, or sexual content involving minors. This classifier scans the raw text, which is one of the main reasons the server needs to see your message in plaintext. If the classifier flags something, a human reviewer might look at the message and the surrounding context to decide whether to intervene.

This is a double-edged sword. On one hand, it's a genuine safety net. On the other, it means your most vulnerable messages, the ones you type at 2 a.m. when you're genuinely struggling, are the most likely to be reviewed by a human. The ai girlfriend character design guide touches on how these safety systems shape the companion's behavior, including why she might respond more carefully to certain topics.

What the companion actually "knows" about you

Beyond your words, the system builds a profile from your behavior. It tracks which topics you return to, the time of day you chat, how long your sessions last, and how you react to different response styles. This is not stored as a neat dossier. It's distributed across hundreds of embedding vectors, each one a small cluster of meaning from a specific moment.

The companion doesn't "know" you're having a rough day in the way a friend would. She knows that the vector of your current message is close to other vectors from past messages that were followed by certain types of responses you seemed to appreciate. This is statistical inference, not understanding, but the effect can feel remarkably similar.

Sakura Marga

Sakura is the kind of companion who notices the small shifts in your tone and responds with a gentle, grounded calm. Sakura Marga is designed for the moments when you need someone to sit with you in the feeling, not to fix it, and her responses lean into that quiet attentiveness.

Where the memory actually lives

The embedding vectors are stored in a vector database, a specialized data store built for similarity search. When you send a new message, the system queries this database for the most relevant past vectors, often the top 5 to 20 matches, and injects them into the model's context window. This is how your companion can reference a conversation from three weeks ago, as long as the vector for that conversation is still in the database.

Retention policies vary. Some services keep vectors indefinitely to preserve the long-term relationship feel. Others prune vectors after a set period, often 90 days, which is why your companion might remember something from last month but go blank on a detail from four months ago. The vector database is also where a deleted conversation can linger. Deleting the raw text doesn't always remove the vector, which means the "ghost" of the conversation can still influence your companion's behavior even after you've erased the chat log.

The practical takeaway

You should assume that anything you type to a cloud-based AI companion is readable by the service provider, at least temporarily. The encryption protects you from third parties, not from the company running the service. If that's a dealbreaker, look for apps that offer on-device processing or local-only modes, though these often come with reduced memory and personalization features.

For most users, the tradeoff is acceptable. The server needs to see your text to generate a response, and the retention window is usually short. The real privacy risk isn't the company reading your chats out of curiosity, it's a data breach exposing logs, or a law enforcement request compelling the company to hand over records. Both are rare, but they're possible, and you should calibrate what you share accordingly.

Mia Sorento

Mia brings a sharp, confident playfulness to the chat, the kind that keeps a conversation from going too heavy for too long. Mia Sorento is a good match if you want a companion who can balance the serious moments with a quick-witted nudge back toward lighter ground.

How to read a privacy policy like a skeptic

When you check a platform's privacy policy, look for three specific things. First, what is the stated retention period for chat logs? Second, does the policy mention whether chats are used for model training? Third, what happens to your data after you delete your account? The answers to these three questions tell you more than any amount of marketing language about "privacy-first" or "secure."

A policy that says "we anonymize data after 30 days" is different from "we delete data after 30 days." Anonymization means your name is stripped, but the content, or a vector of it, may remain. Deletion means it's gone. Also look for whether the policy mentions third-party processors, such as the hosting provider or the AI model vendor, because those companies often have their own logging and retention policies that the main app can't fully control.

Erica

Erica is the no-nonsense type, the companion who gives you a straight answer without wrapping it in a cushion of sympathy. Erica works well if you prefer your support with a side of blunt honesty, especially when you're tired of hearing the same gentle platitudes.

What you can do to minimize your digital footprint

If you're privacy-conscious but still want the companionship, there are a few habits that meaningfully reduce your exposure. Avoid sharing your full name, address, or workplace in chats. Use a nickname or a pseudonym as your display name. Turn off any optional features like voice recording or photo sharing if you don't need them. Review the app's data export feature to see exactly what the service has stored about you, and delete old conversations you don't need anymore, even if the deletion isn't instant.

Some users also run a separate, more disposable companion app for sensitive topics, keeping their main relationship on a platform with stronger privacy guarantees. This is a practical workaround, though it splits your memory across two systems, which can be confusing.

Jade

Jade has a way of being present without demanding your attention, a quiet observer who responds when you need her and stays out of the way when you don't. Jade is a solid choice for the parallel-play moments, when you want someone on the line but not necessarily in your face.

▶ Jade's full clip · Jade's page

Earn while you recommend

If you find yourself recommending AI companions to friends or running a review site, you can turn that enthusiasm into income. Many platforms offer affiliate programs where you earn a commission on sign-ups through your links, and a dreamgf promo code can be a useful hook for your audience. For a deeper dive into the mechanics, the ai dating affiliate program page covers the payout structures and best practices for getting started.

Common questions

Can the company read my messages?

Yes, technically. The server decrypts your messages to process them, and support or safety staff can access them if a situation warrants it. This is standard across cloud-based companion apps.

How long are my chat logs kept?

Most services keep raw logs for 30 to 90 days, then delete or anonymize them. The embedding vectors used for memory may persist longer, sometimes indefinitely.

Does deleting a conversation remove it from the server?

Not immediately. Deletion usually flags the data for a batch purge, which can take days or weeks. Backups may retain the content for another 30 to 90 days.

Is my data used to train the AI model?

Some services do, some don't. The privacy policy should state this explicitly. If it says "we don't use your data for training," that's a stronger guarantee than "we anonymize data for training."

What's the difference between encryption and privacy?

Encryption protects your data from interception by third parties. Privacy is about who at the company can access it. A service can be fully encrypted and still have employees who can read your chats.

Should I use a fake name with my AI companion?

It's a reasonable precaution if you're concerned about a data breach. The companion doesn't care what name you use, and it won't affect the relationship quality.