AI girlfriend vs AI companion: choose the framing before the character

Girlfriend and companion are not two grades of the same product. They are two framings of a fictional conversation, and the one you pick changes which cards are worth opening, which filters help, and what a reply has to do before you call it good. Settling the framing first is cheaper than opening twenty cards and waiting for one of them to feel right.

The framing is a decision about scope

A girlfriend framing narrows the fiction to a single continuing relationship. You are asking one character to hold a shared history, use your name, keep track of what you have told it about your life, and stay inside a romantic register across sessions. Continuity matters more than variety, so a card that opens well but drifts by turn thirty has failed at the thing you came for.

Companion framing widens the scope instead. The relationship may be friendly, mentoring, adversarial, or simply someone to talk at during a late shift, and it does not have to be the only one. Under this framing, running three or four characters for different moods is normal rather than a sign of indecision. What matters is the range of the character and whether it can hold a conversation that has no romantic arc to lean on.

Neither framing is more advanced than the other. They fail differently: girlfriend framing fails on forgotten history, companion framing fails on flatness.

Vendors name these modes differently

Replika’s help material describes its character as having relationship choices such as friend, girlfriend, boyfriend, and mentor, which makes the framing an explicit setting on one persistent character. Nomi’s public AI girlfriend page describes the relationship as a product category with its own feature list. Character.AI documents creator-side surfaces in its help center and leaves the framing to whoever wrote the character. Charmi sits closer to that last shape: the framing lives in the card and in how you open it, not in a mode switch.

None of that says which product suits you. It says where each one expects the decision to be made, which is the only part that affects your next click.

What we observed

Charmi cards carry a tags array alongside an nsfw_level field scored 0 to 3. SFW cards commonly include the literal tag SFW and sit at level 0, so the tag and the numeric field are two views of the same distinction rather than a single toggle. On charmiai.com/explore the filters are gender (all, female, male, nonbinary), a sort control (hot, chats, stars, newest), an adult toggle, and tag filters. The age gate is independent of all of them: the site is 18+ whether or not the adult toggle is on.

Scale forces a filtering decision on you. The imported catalogue is roughly 175,000 cards, and Charmi Chat, this site, lists only the ones a person has reviewed. Those listings are a snapshot with a reviewedAt date, and a fingerprint of each card’s public text is stored so that the sync script refuses to update a card whose name, cover, creator, or context text has changed until someone reviews it again. What you read here is a dated, human-checked subset, not a live mirror of the catalogue.

The hubs on this site describe intent rather than exclusivity. A card listed under a relationship-led hub is not barred from working as a general companion, and the reverse holds too; the hub records what the card was built to do, so treat it as a starting filter instead of a category boundary.

One structural detail matters for the choice in front of you: each character has its own memory store per user, and memories from one character do not leak into another. Running several characters therefore costs you repeated setup, because nothing you told the first one is available to the second.

What each framing changes on the way in

Under girlfriend framing, use the gender filter and start from AI Girlfriend Characters or AI Boyfriend Characters, then commit to one card for a week. State two or three durable facts early, since the per-character memory store starts empty and only fills from what you actually say. A week is long enough for continuity problems to show up and short enough that you have not invested much if they do.

Under companion framing, browse AI Companion Characters with the sort control rather than the gender filter, and accept that you are auditioning voices. Keep the setup light on each one; you will be repeating it.

A test that settles it in twenty minutes

Open one card from each hub and give both the same opening: a small true fact about your day and one question. Reply four times, and resist the urge to write a better opener for the card you already like. Then close the tab, come back later, and ask each one something that depends on the fact you gave it.

If the forgetting bothers you more than the sameness, you want girlfriend framing and should optimise for continuity. If the sameness bothers you more than the forgetting, you want companion framing and should optimise for range. Our practical comparison of relationship-led products covers the product-level version of that decision, and AI companion privacy and healthy-use boundaries covers the limits worth setting before either framing becomes a daily habit.