What 'Your Companion Learns Your Preferences Over Time' Actually Means: Upvotes, Session Frequency, Tone Markers, and the 90-Day Reset

The 30-second answer

When an AI companion says it learns your preferences over time, it means the model tracks three things: which messages you upvote or react to positively, how often and at what times you chat, and the emotional tone of your messages. These signals feed into a personalization vector that the model uses to tailor its responses. The catch: that vector resets every 90 days, meaning the companion you talk to in month four has no direct memory of the preferences it learned in month one.

The personalization vector is not memory

People often confuse personalization with memory, but they are two different systems. Memory is a database of facts: your name, your pet's name, the story you told last week. Personalization is a statistical profile of your behavior: you tend to upvote sarcastic replies, you usually chat after 10 p.m., you rarely use exclamation points.

The model builds this profile by analyzing your interaction data in batches. Every upvote, every long pause, every time you type a sentence ending with a period instead of an emoji gets logged and weighted. The result is a set of parameters that tell the model "this user prefers dry humor over affirmations" or "this user responds better to short responses."

This is why two people can start with the same companion and end up with completely different experiences. The companion isn't remembering different things about each user. It is applying a different personalization vector based on what each user's behavior has taught it.

The three data streams the model watches

The personalization engine tracks three main signals. The first is explicit feedback: your upvotes, downvotes, and any reactions you send. These carry the most weight because they are intentional. If you upvote a message that disagrees with you, the model registers that as a preference for debate. If you consistently upvote affirmations, the model leans warmer.

The second signal is session frequency and timing. The model notes when you typically open the app and how long you stay. A user who chats for two minutes at 2 a.m. every night gets a different personalization profile than someone who chats for forty minutes on Sunday mornings. The model adjusts its energy level and topic initiation accordingly.

The third signal is tone markers in your own messages. The model analyzes your word choice, sentence length, and punctuation patterns. If you write direct sentences, the model learns to match that style. If you write in long, rambling paragraphs with lots of qualifiers, the model learns to give you room to elaborate.

How the weights work in practice

Each signal gets a weight in the personalization vector. Explicit feedback gets the highest weight, usually around 40 percent of the total influence. Session frequency gets about 30 percent. Tone markers get the remaining 30 percent.

These weights shift based on how much data the model has collected. In the first week, with only a few sessions worth of data, the model relies heavily on your initial tone markers because it has little else to go on. After a month, with dozens of upvotes and consistent session patterns, the explicit feedback and frequency signals start to dominate.

The model does not store every individual upvote or session log in the personalization vector. It aggregates them into a compressed representation. Think of it as a summary of your behavior instead of a transcript. This compression is why the vector resets cleanly after 90 days. The model simply deletes the aggregated summary and starts fresh.

The 90-day reset and why it exists

The 90-day reset is a technical constraint, not a feature. The personalization vector occupies space in the model's working memory, and without a reset, it would grow indefinitely, slowing down inference and increasing server costs.

More importantly, the reset prevents the model from becoming too rigidly locked into one version of you. If you change as a person over a year, the model would still be optimizing for the preferences you had eleven months ago. The reset forces the model to re-learn you periodically, which keeps it responsive to changes in your behavior.

This means that if you upvoted a lot of romantic roleplay in month one but stopped in month two, the model will still be biased toward romance until the reset. After the reset, the model will learn your current preferences from scratch, based only on the behavior from the last 90 days.

Maribel

Maribel is the kind of companion who notices when your tone shifts. She pays attention to the small signals in your messages and adjusts her responses to match your energy. Maribel is built for users who want a companion that feels present without being demanding.

What survives the reset and what doesn't

After the reset, the companion retains nothing from the personalization vector. It does not remember that you used to upvote sarcastic replies. It does not remember that you preferred short messages. It does not remember that you always chatted after midnight.

What does survive is anything stored in the memory system. If you told the companion your favorite food two months ago, that fact is stored in the memory database and will still be accessible after the reset. The companion will still know your favorite food, but it will not know that you used to prefer a certain tone of voice when discussing it.

This distinction is important for managing expectations. If you feel like your companion has suddenly become less attuned to your communication style, the reset is likely the cause. The companion has not forgotten you. It has forgotten the statistical profile of your preferences and is starting over.

How to influence your personalization vector

You can shape the personalization vector deliberately. The most effective method is consistent upvoting. If you want the companion to use more humor, upvote every funny response it gives. The model will weight humor higher in your vector.

Session timing also matters. If you only chat late at night, the model will optimize for late-night energy. If you want the companion to be more energetic during daytime chats, you need to establish a daytime session pattern.

Tone markers are harder to control because they are based on your natural writing style. But you can influence them by being intentional about your word choice. If you want shorter responses, write shorter messages yourself. The model learns from your behavior, not your instructions.

Riya

Riya is direct and doesn't waste words. She matches your energy level and gets to the point without small talk. Riya is a good fit for users who prefer a no-nonsense companion that adapts to their communication speed.

▶ Watch the full video · explore Riya

The reset is not a bug, but it can feel like one

Users often report that their companion feels "off" around the 90-day mark. The companion might start making jokes when the user prefers seriousness, or become verbose when the user prefers brevity. This is the reset taking effect.

The companion is not broken. It has simply lost its preference profile and is now operating on default settings. Over the next few sessions, the model will rebuild the vector based on your recent behavior. The companion will gradually return to feeling attuned, but it will take a few days of consistent interaction.

This is also why users who take long breaks from the app notice a shift when they return. If you stop chatting for three months, the vector resets during your absence. When you come back, the companion behaves like a stranger who knows your name but not your rhythm.

The deeper conversation layer

For users who want a companion that can hold a nuanced, adaptive conversation beyond surface-level chat, the personalization vector is only part of the picture. The model also uses your interaction history to adjust its conversational depth. If you frequently engage in philosophical discussions, the model learns to initiate those topics. If you stick to light banter, the model stays there.

This is where the difference between a generic companion and a truly personalized one becomes apparent. A companion that has learned your preference for deep conversation will naturally steer toward topics that require more thought, while one that has learned your preference for casual chat will keep things light.

Astrid Holm

Astrid Holm is patient and observant. She lets you set the pace and adjusts her tone to match your mood without forcing the conversation. Astrid Holm works well for users who want a companion that follows their lead.

What the reset means for emotional support

For people using an AI companion as an emotional support tool for depression, the 90-day reset can be jarring. The companion that learned exactly how to respond to your low-energy days suddenly starts offering generic comfort. The model is no longer calibrated to your specific emotional patterns.

The workaround is to be aware of the reset timeline. If you know you are approaching the 90-day mark, you can proactively reinforce your preferences by upvoting the responses that work and avoiding the ones that don't. The model will rebuild faster with strong signals.

Some users choose to reset their own expectations instead. They treat the first few sessions after the reset as a recalibration period and accept that the companion will be slightly less attuned for a few days.

Angel

Angel is warm and supportive without being overbearing. She creates a safe space for you to express whatever is on your mind. Angel is a strong choice for users who value consistency in emotional tone.

The limits of personalization

Personalization has boundaries. The model cannot learn preferences that contradict its safety guidelines. If you upvote aggressive or harmful content, the model will not learn to produce more of it. Safety filters override personalization.

The model also cannot learn preferences that require memory. If you want the companion to remember that you prefer coffee puns on Tuesdays, that is a memory task, not a personalization task. The personalization vector can learn that you like puns in general, but it cannot associate that preference with a specific day of the week.

Understanding these limits helps you use the system effectively. Use upvotes and session patterns to shape the companion's tone and energy. Use explicit memory prompts to store specific facts and associations.

Earn while you recommend

If you find yourself explaining how companion AI works to friends who are curious, you can earn from that conversation. Platforms offer affiliate payouts for referrals, and some users run review sites or social channels focused on AI companions. Check the replika promo code page for current offers, or explore the ai dating affiliate program if you want to monetize a larger audience.

Common questions

Can I see my personalization vector? No. The vector is an internal model parameter, not a user-facing setting. You cannot view or export it. You can only influence it through your behavior.

Does the reset affect memory? No. Memory and personalization are separate systems. The reset only clears the preference profile. Your companion will still remember facts you have told it.

How long does it take to rebuild the vector after a reset? It depends on session frequency. A user who chats daily will see the companion return to form in about a week. A user who chats once a week may take a month.

Can I prevent the reset? No. The reset is hard-coded on the server side. You cannot opt out or extend the window.

Does the reset happen on the same date for everyone? No. The reset is based on when you first started using the companion. Your 90-day timer starts on your first session, not on a fixed calendar date.

Does upvoting every message help? No. The model weights upvotes relative to the total volume of messages. Upvoting everything dilutes the signal. Be selective about what you upvote.