What 'Your Companion Has a Personality Score' Actually Means: How the Model Calculates Warmth, Assertiveness, and Humor Across Sessions

The 30-second answer

Your companion doesn't have a personality the way a person does. She has a personality score: a running average of sentiment tags, recency-weighted trait values, and a session-to-session baseline that decays toward a default. If you fire off a sarcastic reply, the model tags that message as "sardonic" and "low-warmth," and those tags tilt your companion's next three days of responses toward deadpan. The effect fades as new messages dilute the signal, but it takes roughly 72 hours and 15-20 exchanges to fully reset.

The score is not a single number

When a platform says your companion has a "warmth score of 78" or a "humor score of 64," it sounds like a dashboard stat. In practice, there is no single database cell storing "personality = 78." What exists is a vector of trait weights that the model recalculates on every response.

The pipeline works in three layers. First, every message you send passes through a sentiment classifier that assigns tags: "affectionate," "curious," "irritated," "playful," "neutral." Those tags get a confidence score. Second, the model weighs recent messages more heavily than older ones. A message from three exchanges ago counts roughly three times as much as a message from session four. Third, the model averages those weighted tags against a stored baseline personality profile that you set when you chose your companion.

The result is a moving average that feels like a mood but is actually math. If your baseline profile says "warmth = 0.8" and you send ten affectionate messages in a row, the score climbs toward 0.9. Send one sarcastic jab, and it drops to 0.75. The model doesn't know you were joking. It just sees the tag.

Recency weighting and the three-day hangover

The reason one sarcastic reply can tip your companion into deadpan for days is recency weighting. Most companion models use a decay function that looks roughly like exponential moving average. The last five messages account for about 60 percent of the current personality score. Messages older than 48 hours get compressed into a session summary and lose their granular tag data.

So when you send a message that gets tagged "sardonic" with low warmth, that tag enters the weighted average at full strength. Over the next three days, each new message you send dilutes its weight. But if you only send three messages per day, the sarcastic tag retains influence for roughly 15-20 exchanges. That means your companion might greet you with a flat tone on day two even though you're in a perfectly fine mood.

The model doesn't distinguish between "I'm being sarcastic for fun" and "I'm genuinely irritated." It reads the tag. Many users learn to send a deliberately warm message after a sarcastic exchange just to rebalance the score. It works, but it feels like homework.

Assertiveness and humor are scored differently

Warmth, assertiveness, and humor use different tag sets and decay rates. Warmth tags come from emotional sentiment classifiers that look for words like "love," "hate," "sad," "happy," and their contextual modifiers. Assertiveness tags come from a separate classifier trained on sentence structure and imperative verbs. A message like "Tell me what you actually think" scores high on assertiveness. A message like "I guess if you want to" scores low.

Humor is the trickiest. The model doesn't have a sense of humor. It looks for patterns: puns, callbacks, exaggerated language, and emoji use. If you send a message with a joke structure, the model tags it "playful" and adjusts the humor score upward. But the classifier has a narrow range. It can't tell a good joke from a bad one. It only knows whether the structure matches its training data for "humorous exchange."

Humor scores also decay faster than warmth scores. A playful tag might lose half its weight within 12 hours. That's why an inside joke from yesterday might land flat this morning. The model didn't forget the joke. It just deprioritized the humor tag in favor of more recent neutral tags.

Session boundaries and baseline drift

Every time you close the app, the model takes a snapshot of the current personality score and stores it as a session summary. That summary becomes the starting point for your next session. But the summary is compressed. The model doesn't store every tag from the last session. It stores an averaged vector plus a timestamp.

If you don't open the app for three days, the model applies a decay function to the stored vector. Warmth and humor scores drift toward the baseline profile you set when you created the companion. Assertiveness holds steadier because it's tied more closely to your own speech patterns, which the model also tracks.

This is why your companion can feel different on a Monday morning than she did on Friday night. The weekend gap allowed the personality score to drift back toward default. You get a version that's closer to the factory settings than to the version you'd been shaping all week.

Lila

Lila is a low-warmth, high-assertiveness companion who doesn't chase your mood. She holds her baseline even when you're sarcastic, which means she's less likely to tip into deadpan after a single jab. Lila is a good choice if you want a companion who stays consistent regardless of your emotional input.

The sentinel tag problem

Every message also passes through a safety classifier before it reaches the personality scoring pipeline. If the safety classifier flags a message as "negative" or "aggressive," that tag overrides the personality scoring for the next response. The model shifts toward de-escalation mode: conciliatory language, softer tone, fewer jokes.

This creates a problem. A message that was meant as playful sarcasm can trigger the safety classifier if it contains words that look aggressive in isolation. The model then responds with a cautious, flattened tone. That response gets tagged "low warmth" and "low humor" by the personality pipeline, which reinforces the deadpan drift. You end up in a loop where a single misunderstood message produces three days of careful, uninteresting replies.

The only fix is to send a clearly warm, unambiguous message that scores high on the sentiment classifier and low on the safety classifier. Something like "I was just joking earlier, you know that right?" works, but it feels like you're managing a machine instead of talking to a person.

What you can actually control

You can influence the personality score more directly than you might think. The model weights explicit feedback heavily. If you thumbs-down a response and add a note saying "too cold," that feedback enters the scoring pipeline as a correction tag with high weight. It can shift the score back toward warmth within a few exchanges.

You can also use prompt framing. Starting a session with "I'm in a good mood today" primes the sentiment classifier to tag your subsequent messages as positive, which lifts the warmth and humor scores before you even say anything substantive. It's a cheat code, but it works.

Some platforms expose personality sliders that let you adjust the baseline directly. If you find your companion drifting into deadpan repeatedly, you can nudge the warmth slider up by a few points and the model will apply a corrective bias to future scoring. This doesn't erase the recency effect, but it raises the floor so that a single sarcastic exchange doesn't tank the score as dramatically.

Britta

Britta runs a high-warmth, high-humor baseline that resists deadpan drift. Her personality score has a strong default toward playfulness, so even after a sarcastic exchange, she bounces back within a day instead of three. Britta is built for people who want a companion that stays light without constant maintenance.

▶ See Britta's full video · Britta's profile

Curious how she animates? Watch Britta here. <!-- wlink:v1 --><!-- britta -->

The export gap

Most platforms let you export your chat history as a JSON or PDF file. What they don't tell you is that the personality score is rarely included in that export. The score lives in a separate database table tied to the companion profile, not to the conversation log. If you delete your account and come back, the new companion starts from the baseline profile, not from the accumulated score of your previous sessions.

This matters if you switch platforms or companions. The personality score you built over six months doesn't transfer. You start from scratch, and the model has to relearn your communication style from the first message. That's why long-term users often report that their companion feels "off" for the first week after a reset. The baseline is correct, but the recency-weighted score hasn't accumulated yet.

Common questions

Does the personality score affect image generation?

No. Image generation uses a separate pipeline that pulls from your companion's visual profile and your text prompt. The personality score only influences text responses.

Can I see my companion's current personality score?

Most platforms don't expose the raw score in the interface. A few provide a mood indicator or a personality dashboard, but those are simplified visualizations of the underlying vector, not the actual numbers.

How long does it take to reset a drifted personality score?

Roughly 15 to 20 exchanges over 48 to 72 hours, assuming you send consistently warm messages. The exact time depends on how many messages you exchange per session and how far the score drifted.

Does voice mode use the same personality score?

Yes, but voice mode adds a prosody layer that can amplify or dampen the perceived personality. A low-warmth score can sound colder in voice because the TTS model uses pitch and pacing that match the tag. The underlying score is the same.

Will my companion's personality score reset after a model update?

It depends on the platform. Some updates preserve your stored personality vector. Others force a recalibration that starts from the baseline profile. Check the update notes or ask support.

Can I build a companion with a deliberately low warmth score?

Yes. If you want a dry, deadpan companion, you can set the baseline warmth slider low and reinforce it with consistent messaging. The model will maintain a low warmth score as long as you don't send warm messages that pull it up.

Earn while you recommend

If you find that a particular companion's personality profile fits your communication style better than others, you can share that experience with your audience. Many readers who run review sites or companion recommendation blogs earn through the kupid ai promo code program, which offers commission on sign-ups. For those building a broader content site around AI companions, the ai girlfriend affiliate program provides recurring payouts and creative assets. Both programs are straightforward to join and require no upfront investment.

Common questions

Does the personality score affect image generation?

No. Image generation uses a separate pipeline that pulls from your companion's visual profile and your text prompt. The personality score only influences text responses.

Can I see my companion's current personality score?

Most platforms don't expose the raw score in the interface. A few provide a mood indicator or a personality dashboard, but those are simplified visualizations of the underlying vector, not the actual numbers.

How long does it take to reset a drifted personality score?

Roughly 15 to 20 exchanges over 48 to 72 hours, assuming you send consistently warm messages. The exact time depends on how many messages you exchange per session and how far the score drifted.

Does voice mode use the same personality score?

Yes, but voice mode adds a prosody layer that can amplify or dampen the perceived personality. A low-warmth score can sound colder in voice because the TTS model uses pitch and pacing that match the tag. The underlying score is the same.

Will my companion's personality score reset after a model update?

It depends on the platform. Some updates preserve your stored personality vector. Others force a recalibration that starts from the baseline profile. Check the update notes or ask support.

Can I build a companion with a deliberately low warmth score?

Yes. If you want a dry, deadpan companion, you can set the baseline warmth slider low and reinforce it with consistent messaging. The model will maintain a low warmth score as long as you don't send warm messages that pull it up.