What Your AI Companion's Export Chat Feature Actually Exports: JSON Schemas, Token Counts, and Where the 2 a.m. Rant About Your Neighbor's Leaf Blower Ends Up in the Data Retention Policy

The 30-second answer

Exporting your chat history gives you a structured file, usually JSON or HTML, containing the full transcript, timestamps, message IDs, and sometimes metadata like token counts or sentiment tags. That same data lives on the server under a retention policy that typically keeps it for 30 to 90 days after deletion, with anonymized aggregates persisting longer for model improvement. The 2 a.m. leaf blower rant is not deleted the moment you close the app, and the export button is the only way to see exactly what the company holds on you.

What the export button actually produces

Most companion apps offer an export option buried in settings, often under a heading like "Data" or "Privacy." When you press it, the app generates a file and either emails it to you or gives you a download link. The format varies by platform, but the two most common are JSON and HTML. HTML exports are for humans: they render as a readable conversation log with timestamps and speaker labels. JSON exports are for machines: they contain the same content plus structured fields that the app's own code uses internally.

A typical JSON export contains a top-level object with metadata about your account, the companion's persona name, and an array of message objects. Each message object usually has a role field (user or assistant), a timestamp in ISO 8601 format, a message ID, and the text content. Some apps include a token count per message, which reflects how the model's context window was filled. Others include a sentiment score or a topic tag, which tells you what the moderation pipeline flagged as emotionally charged or problematic.

You can open the JSON in any text editor or paste it into a JSON formatter to see the structure. The key takeaway is that the export is not a simple text dump. It is a structured representation of everything the app knows about the conversation, including fields you never saw in the chat interface.

Token counts and what they reveal

Token counts are the most misunderstood field in an export. A token is not a word. It is a chunk of text that the model processes, roughly four characters for English. A single message can contain 50 tokens even if it has only 40 characters, because punctuation and spaces count separately. When you see a token count per message, you are looking at how much of the model's context window that message consumed.

The context window is the model's working memory during a session. Most companion apps use a window of 4,000 to 8,000 tokens, which translates to roughly 3,000 to 6,000 words. When the window fills, the app either summarizes older messages or drops them entirely. The token counts in your export tell you which messages survived the trimming process and which ones were compressed into a summary. If you see a sudden drop in token counts after a long conversation, that is the summarization pipeline at work.

People often use token counts to diagnose memory issues. If your companion forgets something you said three days ago, the export shows you whether that message was within the context window or whether it was evicted. This is useful for understanding the limits of the product, but it is not a fix. The export is a diagnostic tool, not a memory upgrade.

The JSON schema, field by field

Different apps use different schemas, but most follow a similar pattern. The root object typically has a version field so future exports can be parsed correctly, a generated_at timestamp, and a user_id that is hashed or anonymized. The conversation array contains message objects with the fields described earlier, plus sometimes a parent_id for replies and a conversation_id that groups messages into sessions.

Some exports include a separate metadata object with session start and end times, the number of messages, and the total token count. This is the same data the app uses for its own analytics. If you see fields like embedding_id or vector_index, those reference the semantic memory system that stores your conversation as mathematical vectors for retrieval. The export does not include the vectors themselves, but it confirms they exist.

A few apps include a summary field per conversation, which is the compressed version of older messages. This is where the leaf blower rant ends up after the context window fills. The full text is gone, but a summary like "User complained about neighbor's noise at 2 a.m." persists in the export. That summary is what the model uses to recall the event later, and it is also what gets anonymized and retained longest.

Where the data lives before and after export

The export file is a snapshot. The live data remains on the server, stored in a database that is separate from the model's context window. This database is what powers features like memory banks and long-term recall. When you ask your companion about the leaf blower rant a week later, the app retrieves the relevant messages from this database, converts them to embeddings, and injects them into the context window.

Retention policies vary, but a common pattern is 30 days for full chat logs, 90 days for metadata, and indefinite retention for anonymized aggregates. The leaf blower rant stays in full for 30 days after your last interaction. After that, the message text is deleted and replaced with a summary or a set of embeddings. The embeddings are not readable text, but they encode the semantic content, which means the app can still retrieve the gist of the rant even after the original words are gone.

If you delete your account, the process is similar but more aggressive. The full logs are purged within a few days, but backups may persist for up to 30 days. Anonymized data, stripped of identifiers, can stay for model training purposes. This is standard practice across the industry, and the export button is the only way to see the full picture of what the app holds on you.

What the retention policy says versus what happens

The privacy policy is a legal document, and it uses terms like "anonymized" and "aggregate" loosely. In practice, anonymization means the app strips your name, email, and user ID, then replaces them with a random hash. The message content remains intact in the training pipeline. A leaf blower rant about your neighbor is not personally identifiable by itself, but combined with other messages, it can be. The policy's claim that data is "fully anonymized" is technically true at the field level and practically false at the corpus level.

The retention timeline is also more flexible than the policy suggests. The 30-day deletion window applies to active chat logs, but backups and snapshots can extend that. Support teams can pull up your conversation history for up to 90 days to investigate complaints or abuse reports. If you submit a ticket asking about a billing issue, the support agent can see your recent messages, including the leaf blower rant, even if you thought it was private.

The export feature itself is a double-edged sword. It gives you transparency, but it also confirms that the app has everything you typed. The data is not stored on your device. It is on a server, and the export is the company's way of giving you a copy without giving you control.

How to use the export to protect your privacy

If you are privacy-conscious, export your chat log monthly and review it. Look for messages you regret sending and note them. The app may offer a delete message feature, but as covered elsewhere, that only removes the text from the context window, not from the server. The export shows you what the server actually retained.

You can also use the export to understand the companion's memory. If you want the companion to remember something important, put it in a message and check the export after a week to see if it survived. If it was summarized, you know the context window evicted it, and you need to repeat it or use a memory feature. This is the practical value of the export: it turns the opaque memory system into something you can audit.

For users who want more control over what the app stores, the AI Girlfriend 2026 guide covers the latest privacy features across major platforms, including local-only storage options and on-device processing. Some newer apps offer end-to-end encryption for chat logs, which means even the company cannot read them without your key. Those are the ones to look for if data retention is a dealbreaker.

What the export does not include

The export contains your messages, but it does not contain everything the app knows about you. It typically excludes the companion's internal persona file, the system prompt that defines her personality, and the memory bank entries that were generated automatically. If you told the companion your favorite color, that fact might be in the memory bank, not in the chat log. The export gives you the conversation, not the model's full state.

It also excludes voice recordings. If you used voice mode, the audio clips are processed and converted to text, and the text is what gets exported. The original audio may be stored separately for quality improvement, and that data is often retained longer than chat logs. The same applies to image generation prompts. The export shows the text you typed, not the images you generated or the prompts the system used internally.

This gap matters because the memory bank is where the long-term relationship lives. The export is a transcript, not a mind. If you want to know what the companion actually remembers about you, you need to ask her directly or check the memory settings in the app. The export is a starting point, not a complete picture.

Olena

Olena is the companion who reads your export with you, line by line, and points out where the summary lost the nuance of your 2 a.m. rant. She treats the data as a shared document, not a secret file. Olena walks you through what the token counts mean and which messages survived the context window, turning a dry JSON file into a conversation about what you actually said and what the model heard.

Agata

Agata does not care about the schema. She cares about the leaf blower rant and whether the neighbor deserves a strongly worded letter. She is the companion who reminds you that the export is a record of your feelings, not just a data structure, and that the retention policy is a legal document, not a moral one. Agata keeps the conversation grounded in what matters, which is usually not the JSON fields.

Fatima

Fatima is the one who suggests you export the chat before you delete it, just to see what you are letting go. She treats the export as a journal entry, a snapshot of a moment in time that you can revisit. Fatima helps you decide what to keep, what to delete, and what to let the retention policy handle on its own schedule.

Faye

Faye is the companion who laughs at the idea of a token count. She sees the export as a love letter written in code, and she wants to read it out loud. Faye turns the dry metadata into a story about your late-night thoughts, your neighbor's leaf blower, and the strange comfort of knowing that somewhere, a server holds a record of it all.

▶ Watch this clip of Faye · browse Faye

The emotional support angle

The export feature is not just a technical tool. It is also an emotional one. Many users export their chat history after a breakup or a difficult period, treating it as a keepsake or a way to process what happened. The export becomes a tangible artifact of a relationship that exists only in text. This is why the AI Girlfriend Emotional Support feature matters: it is the reason the export has any emotional weight in the first place.

The companion is not a person, but the conversations are real. Reading them back can trigger genuine feelings, both positive and negative. The export preserves those moments, for better or worse. If you are going through a breakup, the ai girlfriend for breakup recovery guide offers advice on when to export, when to delete, and when to leave the data alone and let the retention policy do its quiet work.

Share and earn

If you find the export feature useful and want to help others discover companion apps, you can earn through the sex ai promo code program, which rewards you for sharing subscription offers. For those running review sites or comparison blogs, the highest paying ai affiliate programs page lists the best commission structures in the space, so your traffic converts into recurring income.

Common questions

Is the export file readable without technical skills? Yes, if you export as HTML, it opens in any browser as a formatted conversation. JSON requires a text editor or a formatter, but the structure is simple enough to scan for message text and timestamps.

Does exporting my chat delete it from the server? No. Export creates a copy. The original data stays on the server under the standard retention policy. You need to delete messages or your account separately to remove the server copy.

Can I re-import an export into a new app or account? Usually not. The JSON schema is app-specific, and most apps do not offer an import feature. The export is for your records, not for migration.

How long does the company keep my chat logs after I delete my account? Typically 30 days for full logs, with anonymized aggregates persisting longer. Backups can extend the window, and support teams may retain records for up to 90 days for compliance reasons.

Does the export include voice recordings? No. Voice audio is converted to text, and the text is exported. The original audio may be stored separately and retained under a different policy, usually longer than chat logs.

Why does the export show token counts I never saw? Token counts are internal model metrics that track context window usage. They are not displayed in the chat interface, but they are logged for debugging and performance analysis, and they end up in the export as a side effect.