AI companion memory is not one thing. A service may use current-session context, saved memories or profile facts, summaries, chat history, local browser storage, or several of these at once. Before sharing anything sensitive, find out what is stored, where it lives, who can access it, whether it may be used to improve a model, and how each layer can be removed.
That is the short answer to the privacy question. The rest of this guide turns it into seven checks you can use before enabling memory or trusting a companion with personal information. The checks are deliberately service-neutral, then applied to Kruthika's current published policy where the policy gives a specific answer.
Key Takeaways
- Memory and chat history can be separate. Deleting one may not delete the other.
- A chat's answer is not a storage audit. Read the settings and privacy policy for servers, providers, logs, backups, and model-improvement use.
- Start with low-stakes information. Do not use a companion chat for passwords, identity documents, financial details, or another person's confidential information.
- Test the exit route before you need it. Check how to clear the saved item, its source chat, exports or shared links, and any retention exceptions.
A five-minute memory check
Run this small test before an intimate or identifying conversation. It can show how the interface behaves, but it cannot prove what a provider, log system, or backup retains.
- Open the memory, privacy, and data-use settings. Note whether they apply to new chats only or to existing data as well.
- Share one harmless preference, such as a fictional character you like.
- Ask what the companion remembers and where that information is shown.
- Change or delete the preference, start a new chat, and see whether it still appears.
- Read the policy's last-updated date and test the deletion path for both the saved item and the source chat.
A result that looks reassuring is only a behavioral check. It does not establish that a record was erased from a server, an inference provider, a backup, or a model-improvement workflow.
1. Can you tell when memory is active?
Answer: You should be able to identify the setting, its scope, and its default state without guessing from the companion's tone.
Look for plain labels such as saved memories, reference chat history, personalization, or temporary chat. These are not interchangeable. OpenAI's current memory documentation, for example, describes saved memories and reference chat history as separate controls and notes that availability can vary by plan, region, platform, and workspace. Its memory FAQ is an example of the kind of product documentation to look for, not a rule for every companion.
If the interface is vague, treat the feature as unclear rather than assuming the worst. Pause before sharing sensitive details and look for the policy, help page, or setting that explains what the toggle actually changes.
2. What is actually remembered?
Answer: Ask the service to name the data layer: current-session context, a saved fact, a summary, chat history, a file, or information held in local storage.
“Memory” is often a product label rather than a technical description. A system may use a short context window while a chat is open, save a preference between sessions, create a summary, or retrieve earlier chats. Those layers can have different locations and deletion paths. A preference such as “I like cricket” may be less sensitive than a detailed account of a relationship or health concern, but no personal detail is automatically harmless. The relevant question is whether the service needs it for the purpose and keeps it for no longer than necessary.
The UK's Information Commissioner's Office recommends checking that personal information is adequate, relevant, and limited to what is needed, and that retention is justified and documented across the AI lifecycle. The ICO's AI data-minimisation guidance is written for organisations, but its test is useful for a reader: what is collected, why, where, and for how long?
3. Can you review and correct saved information?
Answer: A meaningful memory control lets you inspect an item, correct it, and remove it without relying on a guess about what the assistant might remember.
Try the harmless-preference test above, then check the product's actual memory manager. OpenAI documents controls for reviewing, correcting, and deleting remembered information, while also warning that a summary may not show every source or detail the system can reference. That is a useful benchmark, not a promise about another service. Features can differ by plan, region, platform, and account type.
Do not treat a conversational reply as proof of deletion. If a companion says “I forgot,” the response only tells you what it will reveal or use in that moment. It does not independently verify backend storage, logs, backups, third-party processing, or model-training data.
Advertisement
4. Is deletion specific and understandable?
Answer: Separate the source chat, derived memory or summary, account history, exports and shared links, and backups. “Delete my account” is not automatically the same as “delete this memory.”
Before sharing personal details, find clear answers to these five questions:
- What removes the original chat or transcript?
- What removes a saved memory, profile fact, or summary created from it?
- What happens to downloads, exports, feedback, or shared links?
- Are logs or backups kept for a stated period?
- Are there safety, fraud-prevention, legal, or provider exceptions?
OpenAI says that deleting a chat alone may not delete a separate saved memory, and that deleted-memory logs may be retained for up to 30 days for safety and debugging. Google says deleting Gemini Apps activity starts a removal process and describes separate behavior for shared chats, Gems, and temporary retention. These are concrete examples of why deletion needs to be described layer by layer, not a claim that every service follows either policy. OpenAI's memory documentation and Google's Gemini Apps Activity guidance show the level of detail worth seeking.
A 2026 preprint, Tracing Users' Privacy Concerns Across the Lifecycle of a Romantic AI Chatbot Experience, analysed 2,909 privacy-related posts from 79 subreddits collected between November 7, 2024 and November 7, 2025. It studies public discussion and collective sense-making, not company systems or deletion success rates. Its value here is narrower: it illustrates why reversibility and exit controls become more important after intimate disclosures have accumulated.
5. Who can access conversation data?
Answer: The policy should distinguish local storage, network transit, service operators, inference providers, support or safety review, telemetry, and backups.
Kruthika's policy, checked on October 20, 2026, says that chat memory is stored in the browser's local storage until the user clears it. It also says message text is sent over TLS 1.3 to an AI inference provider to generate a reply, while technical telemetry is stored on the service's servers. The policy lists a rolling 90-day schedule for raw technical logs and a rolling 14-day backup window. It says providers process prompts to produce responses and are not authorised to use chat content for their own purposes. Read the current Kruthika privacy policy before relying on those details; the provider cascade and policy can change.
That is not the same as saying a chat is invisible or anonymous. Browser-local memory, encrypted network transit, server telemetry, cookies, optional notifications, and third-party processors are different layers. A good policy names the categories, purposes, locations, providers, retention periods, and user controls so readers can decide what to share.
6. Is chat content used to improve AI models?
Answer: Do not infer the answer from a “private” label. Look for a direct statement about model improvement, analytics, human review, safety investigations, feedback, and opt-out scope.
Published consumer policies differ. OpenAI says personal-account chats and remembered information may be used to improve models when its improvement setting is on. Google documents separate Gemini activity controls and a setting for using certain Live audio, video, and screen-share data to improve services. Anthropic says consumer chats may be used when a user opts in, when conversations are flagged for safety review, or under specified programs; it says Incognito chats are not used to improve Claude. Anthropic's model-training explanation makes the exceptions visible. These examples show why a product-specific answer matters.
Ask whether an opt-out is available, whether it applies retroactively, what happens to feedback or flagged content, and whether a provider receives the text. NIST's AI Risk Management Framework is voluntary organisational guidance for managing AI risks and trustworthiness. It is not a consumer privacy certification, so use it as a governance reference, not as evidence about a particular companion.
7. What should stay out of an AI companion chat?
Answer: Share the minimum needed for the task, and keep secrets and high-consequence identifiers out of the conversation.
- Passwords, one-time codes, account-recovery answers, and private keys.
- Banking details, government ID numbers, tax records, and unredacted documents.
- Health information, precise location, workplace or school identifiers, and intimate details that could harm you if exposed.
- Private photographs or recordings that include you or someone else.
- Another person's confidential information, even if you can see it or have permission to discuss the situation.
No online service is risk-free, including one with useful controls. If you are only testing a companion, use an ordinary topic, a fictional scenario, or a non-identifying preference. A short chat can still create logs or telemetry. Our guide to what “no sign-up” does and does not mean explains why guest access, anonymity, and privacy are different claims.
What Kruthika's policy answers, and what it does not
At the time of this update, the policy gives a more specific picture than a generic “private” label:
- Local memory: chat memory, a name or nickname, language preference, and relationship-stage notes are described as browser-local data that stays until the user clears it.
- Inference: message text is transmitted to the active AI inference provider to produce a reply. The provider list is configurable and may change.
- Telemetry: truncated or anonymised IP information, user-agent and device class, referring page, visited paths, latency, and related technical data are described separately from chat memory.
- Retention: the policy states a rolling 90-day purge for raw technical logs and a rolling 14-day backup window. Aggregated counts may remain.
- Limits: local browser storage does not cover network transit, telemetry, cookies, optional push services, or every third-party service listed in the policy.
This is an informational summary, not legal advice or a guarantee about future processing. Re-check the live policy's “Last Updated” date before sharing sensitive information. The Federal Trade Commission's September 2025 inquiry into seven consumer-facing chatbot companies likewise asked about disclosures, data handling, and use or sharing of conversation information; it was a 6(b) study seeking information, not a finding that every company violated privacy law. Read the FTC inquiry announcement.
How to choose the right privacy article next
This page is about memory layers, saved-memory deletion, and the difference between a visible setting and actual erasure. For the wider set of account, device, and sharing questions, use our AI companion privacy checklist. If you are comparing services before choosing one, our guide to how to choose an AI companion for privacy, memory, and boundaries covers selection criteria rather than repeating this deletion test.
The bottom line
A reasonable minimum standard for evaluating AI companion memory is that the service explains what it stores, makes the relevant control visible, provides a practical deletion path, and discloses retention and model-use exceptions. Those are evaluation criteria, not an industry-wide definition of safe memory.
If you cannot work out what a service remembers or how to remove each layer, keep the conversation non-sensitive. Read the current Kruthika policy before using Kruthika; browser-local memory does not make the entire interaction invisible to network providers or telemetry systems.
FAQs
Does deleting a chat delete saved AI memory?
Not necessarily. A service may keep a saved memory or summary separately from the source chat. OpenAI documents this distinction, and Google's Gemini activity controls show that shared chats, Gems, and temporary retention can have separate behavior. Check the service's deletion instructions for both the source and derived data.
Can AI companion chats be used for model training?
Sometimes, depending on the service, account type, settings, feedback, and safety exceptions. OpenAI, Google, and Anthropic publish different controls. Look for the product's current statement on model improvement, human review, opt-out scope, and whether the choice applies to previously collected content.
Is no-sign-up the same as anonymous?
No. No-sign-up usually means an account is not required. A service may still process network information, cookies, device data, telemetry, or message text sent to an inference provider. Read the privacy policy rather than treating guest access as a promise of anonymity.
How can I test whether a memory was removed?
Delete or change a harmless test preference, start a new chat, and check whether it is used again. This is only a behavioral check. For a stronger answer, also inspect the memory manager, delete the source chat, review the retention policy, and check whether the service describes logs, backups, or provider-side processing.
Sources and further reading
- Kruthika.fun Privacy Policy (publisher policy; re-check the live version).
- OpenAI Memory FAQ (memory layers, deletion, and model-improvement settings).
- Google Gemini Apps Activity: Manage and delete your activity.
- Anthropic: Is my data used for model training?.
- ICO AI data-minimisation guidance.
- NIST AI Risk Management Framework.
- FTC inquiry into AI chatbots acting as companions (September 11, 2025).
- Tracing Users' Privacy Concerns Across the Lifecycle of a Romantic AI Chatbot Experience (2026 preprint; public Reddit discourse study, not a platform audit).
