AI Companion Glossary: The Terms That Actually Matter

The vocabulary in this category is doing a lot of work to keep things vague. Several words describe entirely different systems, and at least one is used to sell a feature that isn’t being delivered. Here’s what the terms actually mean.

Memory and personality

Context window

Working memory: everything the model can see during the current session. It can be very large, which makes one long conversation feel coherent — and it resets when you close the app. Frequently marketed as “memory”, which it isn’t across sessions. Full explanation.

Long-term memory

Storage of past conversations outside the session, plus retrieval when something becomes relevant. The expensive kind, and the one that makes a companion bring something up unprompted a week later. The largest quality gap between platforms.

Retrieval

The search step that finds relevant stored memories and feeds them back to the model. When a companion remembers only after you ask directly, retrieval is the part that’s weak — not storage.

Confabulation

Generating a plausible answer instead of admitting a gap. A companion confidently naming a pet you never mentioned is confabulating. More damaging than forgetting, because it makes everything else it recalls unreliable.

Persona

The character definition: traits, backstory, speaking style. Quality shows in whether the persona holds under pressure or drifts toward agreeing with everything.

Persona drift

The slide from the character you built toward generic agreeableness, usually over a long session. The underlying models are trained to be helpful, and helpfulness erodes a character with edges.

Money

Credits, tokens, coins, gems

The second currency. The subscription buys text; this buys images, voice and sometimes extended memory. Different platforms use different names for the same mechanism. Why this makes advertised prices misleading.

Metered platform

One where core features consume credits rather than being included. Recognisable by a separate top-up page and per-image pricing. The advertised subscription is a floor, not a total.

Free tier

A genuinely usable free level, generally good enough to form an attachment before you hit a limit. That timing is the business model. Still the right place to start.

Billing descriptor

The merchant name that appears on your bank statement. Rarely published in advance, and for many people more important than the price. It’s also how you identify who to contact when a charge won’t stop — see the cancellation guide.

Content and limits

NSFW

A label covering at least four different capabilities: explicit text, explicit image generation, voice, and uncensored roleplay. A platform can offer one and none of the others. What it actually covers.

Uncensored

Marketing language. No platform is unlimited — payment processors, app store policies, model provider terms and law all impose boundaries. The meaningful question is whether the limits are consistent, not where they sit.

Moderation layer

A filter running alongside the model that can block output the model would otherwise produce. A common cause of the same request succeeding one day and failing the next.

Jailbreak

Phrasing intended to get a model past its own restrictions. Widely discussed, generally against terms of service, and often the cause of an account suspension that arrives without explanation.

Technology

LLM

Large language model — the system generating the text. Most companion platforms build on third-party models; a few run their own, which costs more and grants more latitude on content.

Fine-tuning

Further training of a base model on specific data to shift its behaviour. It’s what makes one platform’s conversation feel different from another’s using the same underlying model.

Training opt-out

A setting preventing your conversations from being used to improve the platform’s models. Independent reviews of the category have found these to be rare. Privacy guide.

Retention period

How long a platform keeps your data. A stated duration lets you plan; “as long as necessary” means there’s no fixed limit.

Evaluation

Sexya score

A rating out of one hundred across six weighted axes: natural conversation (20), memory and personalization (20), personality (15), image generation (15), interface and mobile (15), price and value (15). Published only when all six have been measured.

Under evaluation

Our status for a platform that has entered the protocol but not completed it. An untested axis doesn’t score zero — it makes the whole score unpublishable, which is why our cards currently show no ratings. The full protocol.

Reported, not scored

Findings we publish without folding into the hundred points: billing discretion, the cancellation path, account deletion. Kept outside the score deliberately, so no platform earns credit for doing the legal minimum.