AI companion privacy and healthy-use boundaries

AI characters are software, not people, professional care, or crisis resources. They can be useful for fictional conversation and creative play, but they should not become the only place you seek guidance, connection, or help. This guide is general information, not medical or legal advice.

Keep a clear purpose and a clear limit

Before opening a chat, name the purpose: a fictional scene, a writing exercise, or a short diversion. Set a time boundary and notice whether the activity begins to interfere with sleep, work, school, finances, or real relationships. The American Psychological Association’s consumer guide recommends treating chatbots as tools rather than human relationships or professional care, and avoiding major decisions based only on chatbot output.

If you want a fictional premise, choose one deliberately through AI Companion Characters or Character AI Chat. A named purpose makes it easier to stop, switch activities, or contact a real person when that is what you need.

The same discipline applies whether the fictional role is a girlfriend, a boyfriend, or an open-ended roleplay scene. Browsing AI Girlfriend Characters, AI Boyfriend Characters, or AI Roleplay Characters with a purpose already decided is a different activity than opening whichever chat loads fastest and seeing what happens. A decided purpose also gives you a natural stopping point: when the scene you set out to have is finished, close the tab instead of continuing because the interface is still open.

Protect sensitive information by default

Do not enter identifiers such as government numbers, full financial details, passwords, private account credentials, or another person’s confidential information. Treat health, legal, and financial details with the same caution. A conversational interface can feel private even when its policy describes collection, processing, retention, service providers, or other uses.

Read the current policy before you share anything personal. Charmi’s displayed privacy policy says it collects account/profile data, messages, and user-created content; it also describes AI infrastructure providers, de-identified or aggregated improvement where feasible, a support-based training opt-out, and deletion or anonymization timelines. Those are policy disclosures, not guarantees; read the current text and its limits yourself.

Other products differ. For example, Nomi’s privacy policy describes chats, customization, usage, and device data, as well as deletion rights and limits, and urges users not to include personally identifying information. Character.AI’s privacy summary also describes data collection, use, service-provider sharing, and regional rights. Policies and controls can change, so recheck them.

A related but separate question is what a companion product does with the content of your conversation, apart from account-level privacy handling. Some products build a running memory from what you say, which persists differently than a stored transcript: it is fed back into future replies, sometimes long after the original message scrolled out of view. If a product describes a memory feature, look for whether it explains what it keeps, what it discards, and whether you can see the result.

Use the controls that are actually offered

Look for conversation deletion, account deletion, data-access, privacy, and training or personalization controls. Use them where offered, and keep a note of the date and request path if you make a deletion or opt-out request. Do not assume that deleting a local copy means every operational record disappears immediately; policies should explain their own timelines and exceptions.

If you are using an 18+ product, read its current terms and keep adult-use boundaries in view. Charmi’s terms describe an 18+ service and fictional, potentially inaccurate AI output. Do not rely on an AI character to validate a high-stakes decision.

Where a product lets you inspect what it has stored about a conversation, use that view rather than guessing. A control you cannot find is a control you cannot rely on. If a product also publishes information about the people who make its characters, such as a creators page, that is a separate surface from your own privacy controls and does not substitute for reading the privacy policy.

What we observed

We looked at Charmi’s own memory system as a concrete example of what a persistent-memory feature can involve, since general advice about “check what a companion app remembers” is easier to apply with a specific model in view. Charmi stores each memory as a record with display text, a category, the source turn it came from, a confidence score, a locale, timestamps, and a state. Categories are identity, preference, relationship, event, boundary, and promise; states are proposed, accepted, corrected, or deleted. Only accepted or corrected memories are added to a character’s prompt, and a newly proposed memory is shown for review before that happens.

Extraction runs after a turn finishes, not while a reply is still streaming, and it is keyed to the turn so a retry does not create a duplicate. By design, the extraction step excludes transient mood, low-value repetition, unsupported inference, and non-durable sexual detail, so not everything said in a chat is treated as something to remember. In the memory sheet a user can list, add manually, accept a proposed memory, correct its wording, delete it, or export it, and a correction takes effect for the next message without needing a reload. Each character keeps its own memory store per user, so memories do not move from one character’s conversation to another’s. After extraction runs, the chat shows a single quiet notice with a review action rather than surfacing the memory silently. Retrieval into a prompt is capped by plan tier and a token budget, but inspecting and deleting your own memories is not limited by either.

This is one product’s implementation, described to show what to look for elsewhere, not a claim that other companion products work the same way or a guarantee that Charmi’s system will not change.

Maintain human support

The APA’s health advisory distinguishes companion and general chatbots from regulated care and recommends privacy awareness and boundaries. The World Health Organization notes limited evidence and risks when general generative AI is used for emotional or mental-health support.

Prioritize real relationships and professional help when you need support. If you are in immediate danger or considering harming yourself or someone else, contact local emergency services or a crisis service now, or reach out to a trusted person nearby. Do not wait for a chatbot response.

The practical boundary is simple: use fictional AI chat intentionally, share less sensitive information than feels natural, review the current policy and controls, inspect a memory feature rather than take its description on faith, and return to human support for care, safety, and important decisions.