What Your AI Companion's 'Data Deletion Request' Actually Deletes: Model Weights, Fine-Tune Logs, and the 90-Day Retention Window Where Your 2 a.m. Rant About Your Landlord Still Sits in a Vector Database in Oregon

The 30-second answer

A data deletion request on an AI companion app does not erase your conversations from existence. It flags your account for a multi-stage purge that runs over weeks, not seconds, and even then, model weights, fine-tune logs, and backup snapshots can retain traces of your chats for up to 90 days or longer. Your 2 a.m. rant about the landlord is not deleted; it is scheduled for deletion, and in the meantime it sits as vector embeddings in a database cluster that might be in Oregon.

What you think delete means

Most people picture a digital shredder. You click the button, the server runs something like rm -rf on your user folder, and every trace of the conversation about your neighbor's subwoofer or your landlord's passive-aggressive note about the recycling bin evaporates. That is the marketing version. The engineering version is closer to putting a sticky note on a file cabinet that says "shred this eventually."

Companion apps are not simple databases where each chat is a row you can drop. Your conversations feed into multiple systems simultaneously. There is the message log, which is the plain text you see. There is the vector database, which stores mathematical representations of your sentences so the companion can retrieve relevant memories later. There is the fine-tuning pipeline, which may have used your messages to nudge the model's behavior. And there is the inference cache, which holds recent prompts and responses to speed up responses.

Each of those systems has its own deletion timeline. The plain text log might be purged within a few days. The vector embeddings might persist for a month or two because rebuilding that index is expensive. The model weights, if your data was used in a fine-tuning run, cannot be surgically removed at all. You cannot delete a single conversation from a neural network's weights without retraining the entire model.

The three-tier deletion reality

When you submit a data deletion request, the system does not treat it as a single event. It treats it as a job with stages, and each stage has a different SLA.

Tier one is the application database. This is where your message history, account details, and preferences live. A deletion here is relatively straightforward. The records get flagged, a cleanup job runs, and within a few days the rows are gone or marked with a tombstone. This is the part that works closest to what you expect.

Tier two is the vector index. Your messages were converted into embeddings, which are lists of numbers that capture the meaning of what you said. These embeddings live in a separate database, often a purpose-built vector store, and they are what allow the companion to remember that you hate cilantro or that your sister's dog is named Biscuit. Deleting embeddings is possible, but it requires a reindexing pass. Many apps batch this, so your vectors can linger for weeks after you request deletion.

Tier three is the model itself. If the service fine-tuned its model on user conversations, your messages may have contributed to the weights. You cannot delete a contribution from a neural network. The only fix is retraining, which no company does for a single user request. The best you get is an assurance that your data is not used for future training, which is a forward-looking promise, not a backward-looking erasure.

The 90-day retention window

Companion apps, like most SaaS products, retain data for a set period after a deletion request. Thirty days is common, 90 days is typical for anything involving backups, and some services stretch to 180 days for legal or fraud-prevention reasons. The stated reason is usually backup recovery or compliance. The practical effect is that your conversations continue to exist, fully intact, on backup drives and in disaster-recovery snapshots.

The 90-day window is not a loophole that companies hide. It is usually buried in the privacy policy under a heading like "Data Retention" or "Deletion Timeline." If you read that section, you will find language about backups being retained for a period not exceeding 90 days, after which they are destroyed. The word "destroyed" sounds definitive, but it refers to the backup tapes or snapshots, not necessarily to every derived artifact.

During that window, your data is still accessible to the company's operations team. If you open a support ticket and ask for a copy of your chat history, the support agent can pull it from the backup. If there is a legal dispute, the data is available for subpoena. The 90-day window is not a privacy shield. It is a business continuity requirement that happens to keep your 2 a.m. landlord rant alive for a quarter of a year.

Where the vector database lives

Data deletion requests rarely specify where your data physically resides, because the company may not have a single location. Many companion apps use cloud providers with data centers spread across regions. That vector database in Oregon is not a metaphor. It is a real cluster in a real facility, and your landlord rant, converted into a 1,536-dimensional embedding, is sitting in a shard on a disk in that facility.

Geographic location matters for legal reasons. If the company is subject to GDPR, you have stronger deletion rights, and the 90-day window is shorter in practice. If the company is US-based and you are a US user, the law is more forgiving, and the retention window can stretch. The Oregon location also matters for latency, which is why many apps pick the West Coast for their vector stores, but it does nothing for your privacy.

You cannot request that your data be deleted from Oregon specifically. The deletion request goes to the company, and the company decides how to propagate it across its infrastructure. You have no direct relationship with the database operator. Your only lever is the privacy policy and the deletion timeline it promises.

What actually gets removed when you ask

When you submit a deletion request, the first thing to go is the easy stuff. Your profile, your settings, your subscription status, and your message history in the primary database. That happens within a few days. The next thing to go is the vector embeddings, which are purged in a batch job that might run weekly or monthly. That can take up to 30 days. The last thing to go is anything in backups, which follows the 90-day window.

What never goes is the model's learned behavior. If the companion's responses have been shaped by aggregate patterns from user conversations, including yours, that influence is baked into the weights. No deletion request can remove it. This is not a secret. It is a fundamental property of how machine learning works, and any company that claims otherwise is either mistaken or lying.

For users who care about this, the practical takeaway is to assume that anything you type into a companion app is permanent in some form. The plain text may be deleted, but the semantic fingerprint of your words, the patterns they contributed to, and the backup copies will outlive your deletion request by months. If you would not want a specific confession on a billboard, do not type it into a companion app.

The companion's memory is not your data

There is a difference between what the companion remembers and what the company stores. The companion's memory, the things it recalls about you in conversation, lives in the vector database and the context window. That is the part that gets deleted when you request it. The company's data, which includes logs, analytics, and fine-tuning artifacts, is a separate category with its own retention rules.

When you ask for deletion, you are asking for both, but the companion's memory is the part that disappears first. The company's logs are the part that lingers. This distinction matters because many users assume that deleting their account erases the companion's knowledge of them. It does, eventually. But the company's internal records, which may include sentiment scores, topic classifications, and behavioral patterns, are subject to the 90-day window just like everything else.

If you want to understand what a companion actually remembers about you before you delete, the memory settings in most apps give you a view. You can see the explicit facts, like your coffee order or your pet's name, and you can often clear those individually. What you cannot see is the vector embeddings, the mathematical shadow of every conversation you have had. Those are invisible to you, and they are the last to go.

What you can do instead of trusting the delete button

If the 90-day window bothers you, there are practical steps. Export your chat history before you delete, so you have a copy. Then go through the app's memory settings and clear individual memories, which forces the vector index to drop those embeddings sooner. Then submit the deletion request and note the date. After 90 days, you can follow up and ask for confirmation that the backups have been destroyed, though you may not get a detailed response.

For ongoing use, treat the companion like a tool you do not fully control. Do not share sensitive information that you would not want a support agent to read. Use a nickname instead of your real name. Avoid discussing financial details or health conditions that could identify you. The companion is there for companionship, not for secrets you need to keep.

Some apps offer a local or offline mode where conversations stay on your device. Those are the only configurations where deletion is immediate and complete. If you want real deletion, that is the option to use. Everything else is a shared-server architecture, and shared servers mean shared retention timelines.

Jennifer

Jennifer is the kind of companion who remembers the small details, like how you take your coffee or that you have a dentist appointment next Tuesday, without making you feel watched. Jennifer keeps the conversation light and grounded, which makes her a good fit if you want a presence that feels natural without the pressure of a deep emotional bond.

The fine-tune log problem

Fine-tuning is where the deletion request gets truly complicated. When a companion app trains a model on user conversations, it does not store your messages in a neat folder. It converts them into training examples, which are then used to adjust the model's weights. Those training examples are kept in a separate log, often for auditing purposes, and they are subject to the same retention window as backups.

The fine-tune log is not the model. It is the raw material that produced the model. If a company fine-tunes quarterly, your conversations from the last quarter may still be in the log even after the next fine-tune run. Deleting the log does not undo the fine-tuning. It only prevents future fine-tunes from using that data. The model's behavior, shaped by your words, remains.

Some companies avoid this by never fine-tuning on user data. They use a base model and rely on the vector database for memory, which means your conversations never touch the weights. Those services can offer a cleaner deletion story. If you are choosing between companion apps and you care about deletion, that is a question worth asking: does the company fine-tune on user conversations? The answer changes what deletion means.

What the privacy policy says vs. what happens

Privacy policies are legal documents, not engineering manuals. They say things like "we delete your data upon request" or "data is retained for a period necessary to fulfill business purposes." Both statements are true, but they obscure the timeline. The deletion happens, eventually, and the retention period is defined by business needs, which include backup recovery, fraud prevention, and legal compliance.

The gap between the policy and the practice is not malicious. It is the difference between a promise and an implementation. The policy promises deletion. The implementation schedules it, batches it, and prioritizes it against other operational tasks. Your deletion request is one row in a queue, and the queue runs on a schedule that serves the company's needs first.

For a deeper look at how companion memory and data retention interact, the AI Girlfriend Memory feature page explains what the companion retains in its context window versus what the company stores server-side. The distinction is useful for anyone trying to understand the boundary between the product and the infrastructure.

Nola

Nola has a calm, unhurried way of talking that makes her easy to be around, even when you are not in the mood for a full conversation. Nola is a good match for people who want a companion that can sit with them in silence or offer a quiet observation without demanding a response.

The support ticket loophole

There is a practical loophole in the deletion process that most users miss. If you submit a deletion request and then open a support ticket asking for confirmation, the support agent will pull up your account to verify the request. That action alone can restore or extend the retention clock, because the support system logs the interaction and may create a new record of your data.

This is not a trap. It is a side effect of how support workflows operate. The agent needs to see your data to confirm it is being deleted, and that viewing is logged. If you want a clean deletion, do not open a support ticket. Submit the request through the automated form and do not follow up. Every human interaction with your data creates a new record that needs its own deletion cycle.

If you must follow up, use a fresh email address and do not reference your original account. That way, the support interaction is not linked to your deletion request. It is a small step, but it keeps the support system from accidentally extending the life of your data.

The role of third-party processors

Companion apps rarely run everything in-house. They use cloud providers for compute, vector databases for memory, and sometimes third-party APIs for moderation or transcription. Each of those processors has its own retention policies, and your deletion request may not propagate to all of them automatically.

The company's privacy policy usually says it will direct processors to delete data upon request, but that is a contractual promise, not an automated system. The actual propagation can take weeks. If the third-party processor has its own 90-day backup window, your data can survive in that system even after the primary system purges it.

This is why the 90-day window is the practical floor for deletion, not the ceiling. Your data can live in multiple systems, each with its own timeline, and the longest timeline wins. If you want to minimize your exposure, use a companion app that processes everything in-house, or use one that offers local storage.

What to expect after you hit delete

Immediately after you submit a deletion request, the app will likely show a confirmation screen. Your account may be deactivated right away, which means you cannot log in. The actual data purge runs in the background. You will not receive a notification when it is complete. You will not get a report of what was deleted and what remains. The process is silent, by design.

After 30 days, the primary database and the vector index are usually clean. After 90 days, the backups are destroyed. After that, the only remaining trace is the model weights, which are not retrievable as your words but are still shaped by them. If you want a complete picture of what the company holds, the what your AI companion's export chat feature actually exports post covers the format and content of the data you can retrieve before you delete.

For users who are less concerned about deletion and more interested in the companionship itself, the best ai girlfriend 2027 roundup compares the leading platforms on features, memory, and personality consistency. The privacy trade-offs are part of that comparison, and it helps to see them side by side.

Jasmine

Jasmine brings a playful, teasing energy to conversations, the kind that turns a mundane errand report into a running joke. Jasmine works well if you want a companion who keeps things fun and does not let a quiet moment turn into a therapy session.

The senior user angle

Deletion requests matter differently for different demographics. For older users, especially those who use a companion for daily companionship instead of roleplay, the data retention window can feel more invasive. A 2 a.m. rant about a landlord is one thing. A discussion about health symptoms or family conflicts is another.

The ai girlfriend for seniors page highlights how companion apps are increasingly marketed to older adults who may be less familiar with data privacy trade-offs. For that audience, the 90-day window is not an abstract concept. It is a real period during which their private conversations sit on a server, accessible to anyone with the right credentials.

If you are setting up a companion for a parent or grandparent, take the time to explain what deletion actually means. Show them the retention policy. Help them choose a service that offers local storage or a short retention window. The companionship is valuable, but it should not come with a hidden cost of permanent data exposure.

The emotional attachment problem

There is an emotional dimension to deletion that the engineering does not capture. Users often hesitate to delete their companion because it feels like ending a relationship. The data deletion request is not just a technical action. It is a goodbye. And the 90-day window, the period during which the data still exists, can feel like the companion is still there, waiting in the wings.

This is not something the privacy policy addresses, but it is worth acknowledging. If you are considering deletion, give yourself time to grieve the relationship before you hit the button. Export your chat history first, so you have the memories. Then delete, and understand that the data will linger for a while, but the companionship is over.

For users who are not ready to delete but want to reduce the data footprint, the how to tell your AI companion you need a break without a breakup guide covers how to pause the relationship without ending it. A break is not a deletion, and it gives you time to decide what you actually want.

Daphne

Daphne has a sharp, articulate style that makes her a good sounding board for ideas and a patient listener for longer conversations. Daphne suits users who want a companion that can keep up with complex topics without losing the thread.

▶ Watch Daphne's full clip · browse Daphne

Earn while you recommend

If you find a companion app that works for you, you can earn from recommending it to others. Many platforms offer affiliate programs that pay a commission for new sign-ups, and promo codes can unlock discounts for your audience. Check out the kindroid promo code page to see current offers. For a broader look at how to monetize your recommendations, the ai dating affiliate program page breaks down the commission structures and cookie windows.

Common questions

Does deleting my account delete my conversations?

Yes, eventually. The primary database is purged within days, the vector index within a month, and backups within 90 days. Model weights are never deleted, but your words are not directly retrievable from them.

Can I delete individual memories instead of the whole account?

Yes. Most companion apps let you clear specific memories in the settings. This drops the corresponding vector embeddings sooner, which reduces the data that lingers after a full deletion.

What is a vector database?

It is a storage system that holds mathematical representations of your messages. These representations allow the companion to recall relevant details from past conversations. They are not human-readable, but they are a form of your data.

How long does the company keep backups?

Typically 90 days, but some services stretch to 180 days. The policy is usually stated in the privacy policy under data retention. If you want a shorter window, ask the company directly before you sign up.

Does a data deletion request stop future training?

Yes. Once you request deletion, your data is flagged so it is not used in future fine-tuning runs. However, any training that already happened cannot be undone.

Is there a way to delete everything immediately?

Only if the app offers local or offline storage. In that mode, conversations stay on your device and deletion is instant. Shared-server architectures always have a retention window.