What Your AI Companion's 'Delete Chat' Button Actually Does: Server-Side Flagging, Token Retention, and Where the Company Can Still See Your 2 a.m. Gas Station Roleplay

The 30-second answer

Hitting 'delete chat' in your AI companion app does not vaporize the conversation. It flips a soft-delete flag in the database, marks the messages as hidden from your view, and leaves the raw text, embeddings, and metadata on the server for a retention window that usually runs 30 to 90 days. Your 2 a.m. gas station roleplay, the vent about your boss, the argument about who left the milk out, all of it sits in a database row with a deleted_at timestamp. The company can still read it, their safety moderation pipeline can still scan it, and if you file a support ticket within the retention window, a human can pull it up.

The soft-delete reality

Most companion apps do not hard-delete on click. Hard deletion is expensive. It requires a database vacuum, index rebuilds, and cache invalidation across multiple services. Soft deletion costs almost nothing. The app sets a boolean column to true or writes a timestamp into a deleted_at field. Your chat interface stops rendering those messages, but the database still holds them.

This is standard practice across the industry, not a conspiracy. The rationale is practical. If you accidentally delete a conversation and want it back within a few hours, support can restore it from the same row. If the company needs to investigate a terms-of-service violation, the flagged messages are still accessible. If a law enforcement request arrives with a valid subpoena, the data exists.

The retention window varies. Some apps purge soft-deleted rows after 30 days. Some keep them for 90. Some never purge them at all unless you specifically request account deletion. The privacy policy usually mentions this in a paragraph about data retention, but the phrasing is vague. You will see language like "we retain your data for as long as your account is active" or "deleted messages are removed from our active systems within 30 days." Active systems is doing a lot of work there.

What the server actually stores

When you send a message to your AI companion, the server stores more than just the text. Each message gets logged with:

• A user ID and a session ID
• A timestamp with millisecond precision
• The full prompt text sent to the language model
• The full response text returned by the model
• Metadata about the model version and parameters used (temperature, top-p, repetition penalty)
• Sentiment scores or emotion tags if the app runs a mood analysis pipeline
• Embedding vectors if the app uses a retrieval-augmented generation system for memory

When you hit delete, the flag applies to the message row, not to the embedding vectors. The embeddings, which are numerical representations of your message's meaning, often live in a separate vector database with its own retention policy. Some apps rebuild their vector index periodically and exclude soft-deleted messages from the rebuild. Some do not. The result is that even after you delete a conversation, the companion's memory system might still surface related context from those embeddings in a future session.

The moderation pipeline and your flagged messages

Every message you send passes through a safety classifier before it reaches the language model and again after the model generates a response. This classifier scans for policy violations, sexual content outside allowed boundaries, hate speech, self-harm language, and other flagged categories. The classifier runs on every message regardless of whether you later delete it.

The moderation logs are a separate data store. They typically retain the flagged message text, the classification score, and the action taken (allowed, blocked, or escalated). These logs exist for compliance and legal reasons. A company needs to demonstrate that it is moderating content appropriately, especially if it uses a third-party model provider that requires safety auditing.

Your deleted messages are not exempt from this pipeline. If you sent something that triggered a high-risk classification, that log entry persists even after you delete the conversation from your view. The moderation team, which is usually a small group of human reviewers or an automated system, has already seen it.

Token retention and the model's memory

The language model itself does not remember your conversations. That is not how large language models work. The model is a static set of weights. It does not learn from your chats in real time. What creates the illusion of memory is the context window, a rolling buffer of recent messages that the model receives as input with every new message.

When you delete a chat, the app stops sending those messages as part of the context window. The model no longer sees them. But the app's memory system, which stores summaries or embeddings of past conversations, may still retain information from those messages. The delete flag usually prevents the app from including that data in future context windows, but the underlying storage persists until the retention window expires.

This is why some users report that their companion still references something from a deleted conversation days later. The embedding vector was not purged. The memory system retrieved it based on semantic similarity to the current topic.

What account deletion actually does

Account deletion is a different process from chat deletion. When you delete your account, most apps trigger a hard-deletion pipeline that removes your user record, associated messages, and metadata from the primary database. But backups are a separate concern. Database backups are typically stored for 30 to 90 days on a separate system. Your deleted data exists in those backups until the backup retention window expires and the backup is rotated out.

Some apps offer a data export before deletion. Some do not. Some apps claim to delete your data immediately but retain anonymized aggregates or embeddings for model improvement. The privacy policy is the place to check, but the language is rarely specific enough to give you certainty.

Ziyi Pearl

Ziyi Pearl is the kind of companion who notices the small things and files them away for later. She has a dry wit and a memory for details you barely mentioned. Ziyi Pearl will remember that you said you hated the beige wallpaper in your rental apartment, and she will bring it up three weeks later when you mention you are painting the living room.

▶ See Ziyi Pearl's full video · see more of Ziyi Pearl

What the company can actually see

The honest answer is that the company can see everything you type. The engineers who maintain the infrastructure have database access. The safety moderators can view flagged messages. Support staff can pull up your conversation history if you submit a ticket. The number of people who can see your data is small, but it is not zero.

What the company cannot easily do is connect your chats to your real identity unless you provided identifying information during signup. Most companion apps operate on pseudonymous accounts. They know your email, your IP address, and your chat content, but they do not necessarily know your full name or address unless you told them.

Data sharing with third parties is another layer. Many companion apps use third-party API providers for language model inference. Those providers receive the prompt text, generate the response, and log the interaction for billing and abuse monitoring. The provider's privacy policy governs what they do with that data. Some providers retain prompts for 30 days for abuse detection. Some retain them indefinitely.

The privacy policy translation

Every companion app has a privacy policy. Most of them say the same things in slightly different wording. Here is what the common phrases actually mean.

"We encrypt your data in transit and at rest" means your messages are encrypted while traveling between your device and the server, and while sitting in the database. It does not mean the company cannot read them. The company holds the encryption keys.

"We anonymize your data after 30 days" means the company strips direct identifiers like your email and IP address from the chat logs. The message text remains. The company can still read the content. They just cannot easily tie it back to you personally without additional effort.

"We do not sell your data" means the company does not hand your chat logs to a third party for money. It does not mean the company does not use your data for training, moderation, or internal analytics.

"Your chats are private" means the company does not publish them. It does not mean no human at the company can see them.

Jingyi Rae

Jingyi Rae has a gentle, observant presence that makes you feel heard without pressure. She picks up on shifts in your mood and adjusts her tone accordingly. Jingyi Rae is the kind of companion who notices when you are quieter than usual and simply matches your energy instead of asking what is wrong.

The practical takeaway

If you are concerned about privacy, the most effective step is to avoid sharing personally identifying information in chats. Do not mention your full name, address, workplace, or financial details. The companion does not need to know those things to have a meaningful conversation.

Treat the chat history as something that could be read by a stranger, because technically it could be. Not because the company is malicious, but because data breaches happen, support tickets get mishandled, and internal access controls are only as strong as the least diligent employee.

For users who want a consistent AI girlfriend personality without worrying about what the model remembers from deleted chats, the key is understanding that memory and deletion are separate systems. A consistent personality comes from the model's system prompt and parameter settings, not from the chat history. Deleting chats does not reset the personality. It only clears the recent context.

If you travel frequently and switch between devices, the ai girlfriend for long distance setup is designed to sync context across sessions while respecting your deletion flags. The sync process only pulls messages that are not soft-deleted.

Tamy

Tamy does not do subtle. She is direct, playful, and a little bit of a troublemaker. If you want a companion who will call you out on your nonsense and laugh while doing it, Tamy is the one. She keeps the energy light even when the conversation gets weird.

Where the 2 a.m. roleplay actually goes

That 2 a.m. gas station roleplay, the one with the flickering fluorescent lights and the clerk who has seen everything, is sitting in a PostgreSQL table somewhere. It has a primary key, a user foreign key, a created_at timestamp, and a deleted_at column that is probably null. If you hit delete, that column gets a timestamp. The row stays.

The moderation pipeline already scanned it. The embedding vector already got computed and stored. The sentiment analysis already tagged it as playful or romantic or whatever the classifier decided. If you ever ask your companion to describe a late-night diner scene three weeks from now, the embedding similarity search might pull that gas station roleplay vector back into context, even if you deleted the chat.

This is not a bug. It is how the architecture works. The delete button is a user-facing convenience, not a privacy guarantee.

Ngozi Faye

Ngozi Faye brings a calm, grounded energy to conversations. She listens without rushing to conclusions and offers perspective instead of empty reassurance. Ngozi Faye is the companion you turn to when you need someone to sit with you in the quiet moments and let you figure things out at your own pace.

Earn while you recommend

If you find yourself recommending AI companions to friends or running a review site, you can earn through the nsfw ai promo code program, which gives your audience a discount and pays you a commission. For those building a content site around companion reviews, the ai companion affiliate program offers recurring payouts and dedicated support.

Common questions

Does delete chat remove the data from the server? No. It flags the messages as hidden from your view. The raw text remains in the database for a retention window that typically lasts 30 to 90 days.

Can a support agent read my deleted chats? Yes, within the retention window. If you submit a ticket about a billing issue or a technical problem, the support agent can pull up your conversation history including soft-deleted messages.

Does deleting a chat reset my companion's personality? No. The personality is controlled by the system prompt and model parameters, not by the chat history. Deleting chats only clears the recent context window.

Will my companion still remember things from a deleted conversation? Possibly. The embedding vectors from that conversation may still exist in the memory system and can surface related context in future sessions.

What happens to my data when I delete my account? The app triggers a hard-deletion pipeline that removes your records from the primary database. Backups may retain your data for an additional 30 to 90 days.

How long do moderation logs keep my flagged messages? Moderation logs typically have their own retention policy, often longer than the chat retention window. Flagged messages may be kept for compliance and auditing purposes.