What "Memory as a Default" Means for Your Privacy Now That Every Major AI Assistant Retains Conversations

Three major AI assistants now remember your conversations by default. The defaults, the retention windows, and the opt-out mechanics differ enough across all three that even experienced users get confused. Privacy is not a single toggle. It is at least three separate systems, and most coverage conflates them. This piece breaks down what "memory on" actually means at each provider, what turning it off does and does not stop, and where the industry's defaults leave you.

Key Takeaways

• "Memory," "retention," and "training" are three independent consent surfaces. Turning off one does not turn off the others. Most privacy guides treat them as a single switch, which is wrong.
• Every major assistant's "temporary" or "incognito" mode still retains data for some window: 30 days, 72 hours, or organization-dependent. Temporary does not mean zero.
• The safest defaults (memory off, admin-controlled retention) are reserved for enterprise and education tiers. Consumer tiers get memory on, with the burden on you to find the opt-out.
• Deletion is never instant. Deleted chats, deleted memories, even deleted accounts persist on provider servers for days to weeks before actual removal.
• If you build or use a product that layers on top of these providers, you inherit their retention floor. "Zero-retention" as a product claim is almost always false for anything that touches a frontier model at inference time.

Why Does the Trade Press Keep Getting This Wrong?

Because "training" and "retention" sound like the same thing, and they are not. A provider can retain your conversation for safety review, abuse monitoring, or debugging for weeks or months without ever feeding it into a training run. A Windows Forum comparison piece made this point explicitly: a lot of privacy advice fails because it treats "training" as identical to "retention." When you toggle off "improve the model for everyone," you have opted out of training. You have not opted out of retention. The chat still sits on a server somewhere, often for longer than you would guess.

And then there is personalization memory, which is a third system entirely. It is the part where the assistant remembers that you prefer metric units, that your dog's name is Biscuit, that you are a backend engineer who works in Go. This layer persists across sessions and is distinct from both the raw chat log (retention) and the model-training pipeline. Three systems, three defaults, three opt-out surfaces. One toggle does not cover them.

How Does ChatGPT Handle Memory by Default?

ChatGPT's memory system has two layers: an explicit, user-editable list of "saved memories" and an implicit "reference chat history" feature that recalls patterns from past conversations. The reference history feature launched in April 2025. Both layers are on by default for Free, Plus, Pro, and Team plans. Enterprise and Edu tiers flip the default: memory is off unless a tenant admin opts in.

That split is worth pausing on. The same company ships two opposite defaults depending on whether you pay for a business plan. The most protective default is reserved for the customer with the most negotiating power.

Deletion is not instant. Every form of deletion in ChatGPT carries a 30-day window where data still exists on OpenAI's servers for safety monitoring. That applies to deleted chats, expired temporary chats, deleted memories, deleted files, and full account deletion. Thirty days is not a bug; it is the documented design. If your threat model cares about data at rest on a third-party server, "delete" buys you a 30-day countdown, not an immediate wipe.

What Does Gemini Retain, and for How Long?

The default retention window for Gemini activity is 18 months, adjustable via auto-delete settings. Human-reviewed conversations can be kept for up to three years. The cross-app data access layer, now branded "Personal Intelligence," pulls from Gmail, Drive, Maps, and other services. The provider has stated this data is used only to power individual-user features rather than to train public models, but the default retention window is still 18 months of your activity sitting on their infrastructure.

When Gemini's personalized-response feature is available, it is enabled by default. You must manually navigate to Settings, then Personal context, to stop the assistant from remembering your details. The temporary chat mode retains content for up to 72 hours before deletion, which is better than 30 days but still not zero.

There is also active legal scrutiny. A lawsuit, Thele v. Google LLC in the Northern District of California, alleges that the existing "Smart Features" toggle was used as a backdoor to enable Gemini's deeper data access without a clear second consent prompt. As of March 2026 the case was in discovery. The outcome matters for how "default on" consent mechanisms get treated legally going forward.

What Is Copilot Doing with Your Conversations?

For Microsoft's consumer Copilot, personalization is turned on by default wherever it is available. Conversations are saved by default, letting you view and access past sessions. Microsoft retains Copilot conversation history for 18 months by default.

And then there is the secondary toggle. Investigative reporting in February 2026 found that Microsoft had quietly added a default-on setting permitting Copilot to use data from Bing, MSN, Edge, and other Microsoft products the user has interacted with. Reporters found this toggle turned on by default for many users, with no prominent disclosure at the time it was enabled.

On the enterprise side, Microsoft 365 Copilot documentation confirms that no action is required to turn on Copilot memory. Tenant admins only need to act if they want to turn it off. "Enhanced personalization," the control governing memory use, is on by default. The enterprise customer at least has an admin who can flip the switch. The consumer does not.

Is "Temporary Chat" Actually Temporary?

No. Not at any of the three providers.

ChatGPT's temporary chat mode still results in data persisting for up to 30 days on their servers. Gemini's temporary chat retains content for up to 72 hours. Copilot's temporary chat retention depends on the organization's data retention policy and may still be accessible to IT admins during that period.

"Temporary" is a UX label, not an architectural guarantee. The branding gives users confidence that the conversation vanishes when they close the tab. It does not. If your concern is that sensitive data should never touch a provider's disk, temporary mode at any of these three does not solve it. It shortens the window. It does not eliminate it.

Why Do Enterprise and Consumer Defaults Diverge?

Because enterprise customers negotiate contracts, and contracts specify data handling terms that regulators and procurement teams actually read. Consumer users accept a terms-of-service update they never opened.

The result is a two-tier system. OpenAI defaults memory off for Enterprise and Edu. Microsoft gives tenant admins the kill switch but leaves consumer Copilot on by default. Consumer Gemini ships with personalization on by default. In every case, the safer default costs more or requires organizational infrastructure to access.

This is not a conspiracy. It is an incentive structure. Memory makes the product stickier. Stickier products retain users. Retained users generate more data. More data improves personalization. The loop is obvious, and breaking it for consumers costs engagement metrics. Enterprise customers break it because their legal departments require it.

What Does "Opt Out" Actually Turn Off?

Less than you think, and it varies by provider. Here is the real matrix.

If you turn off memory in ChatGPT, you stop the assistant from saving new explicit memories and from referencing chat history across sessions. You do not stop OpenAI from retaining your chats for up to 30 days for safety monitoring. You also need a separate toggle to opt out of model training.

If you turn off personal context in Gemini, you stop personalized responses. You do not automatically reduce the 18-month activity retention window. That requires going to your activity controls and adjusting auto-delete settings separately.

If you turn off personalization in Copilot, you stop the assistant from using your profile to tailor responses. You may still need to separately disable the secondary toggle that permits use of data from Bing, MSN, and Edge.

Three providers. Three different "off" switches. None of them turns off retention entirely. The Windows Forum comparison piece was right: the defaults, names, and scopes vary enough to confuse even experienced users.

Why Is "Zero-Retention" Almost Always a False Claim for LLM Products?

Because any product that sends a request to a frontier model at inference time inherits that provider's retention floor. You can encrypt everything on your side, store nothing in your own database, and still have data sitting on someone else's servers for 30 days or 72 hours or 18 months because the inference call landed there.

We know this because we build one of these products. Selina's memory is encrypted at rest, but it is NOT end-to-end encrypted and NOT zero-knowledge encrypted, because a slice of each request reaches a frontier provider at inference. We route requests through a stack of frontier models, routed per task. That means the provider sees enough of the prompt to generate a response. Claiming zero-retention for the memory layer would be dishonest given that architectural reality. Non-content operational metadata is kept for a short retention window. We are explicit about this because the alternative is saying something that sounds good and is not true.

Files and transfers through SelinaSEND are a different story. Those can be and are end-to-end encrypted, zero-knowledge. The distinction matters: file transfer is a pipeline we control entirely. Inference is a pipeline we share with a provider. Different architectures, different honest claims.

Any product that tells you it has "zero retention" while also calling a third-party model API is either misinformed about its own supply chain or counting on you not to ask. The question to ask is: what is the retention floor of every provider in your inference path? If the answer is "we don't know" or "we'd rather not say," that tells you enough.

What Should Privacy-Conscious Users Actually Do?

Start by treating "memory," "retention," and "training" as three separate problems, because they are.

For memory (personalization across sessions): find the specific toggle at each provider. ChatGPT: Settings > Personalization > Memory. Gemini: Settings > Personal context. Copilot: Privacy controls in settings. Turn them off if you do not want cross-session recall.

For training: find the model-improvement or "help improve" toggle. This is separate from memory at all three providers. Turning off memory does not turn off training opt-in. Check both.

For retention: understand that you cannot reduce it below the provider's floor. You can shorten Gemini's window from 18 months to 3 months via auto-delete settings. You cannot make ChatGPT delete faster than 30 days. You cannot control Copilot's retention if your organization sets a policy. Adjust what you can. Accept that a floor exists.

And then make input decisions accordingly. If a piece of information should not sit on a third-party server for 30 days, do not type it into any of these assistants, regardless of what mode you are in. The most reliable privacy control is still what you choose to share in the first place.

What Does This Mean for Product Builders?

If you are building on top of these models, you inherit their retention characteristics whether you acknowledge it or not. Your privacy policy needs to account for the downstream provider's data handling, not just your own. Your users deserve to know the full chain of custody, including the parts you do not control.

The Wall Street Journal flagged that major AI assistants now retain details users shared casually, sometimes months earlier, without prominent disclosure at the time of input. If you are building a product that adds memory on top of a provider that also retains data, your user's information now lives in two places with two different retention policies. That is a real compliance surface, not a theoretical one.

The honest path is to state what you encrypt, what you do not, what your providers retain, and for how long. We made this choice at Selina: memory encrypted at rest, NOT end-to-end encrypted, short retention window for operational metadata, zero-knowledge encryption for files and transfers only. The account is protected. Content is encrypted. We do not claim what we cannot architecturally guarantee. It costs us marketing copy. It saves us from lying.

Where Is This Going?

Memory as a default is not going away. xAI has added a comparable memory feature to its assistant, available in beta outside the EU and UK. Every major assistant is converging on the same bet: that users want continuity across sessions, and that the engagement gains from remembering outweigh the privacy costs. The trend is expanding across the industry, not contracting.

Regulation will eventually catch up. The Thele v. Google case may set precedent for whether a pre-existing toggle can serve as consent for a fundamentally new data use. GDPR enforcement around cross-session memory in AI assistants is still early. The EU's exclusion from some memory features (like xAI's) suggests providers already expect friction.

In the meantime, the burden sits with you. Read the settings. Count the toggles. Understand that "off" means something different at each provider, and that none of them means "gone." The defaults are chosen to benefit the platform. Your job is to decide whether those defaults also benefit you.

If you want an assistant that encrypts content at rest, applies zero-knowledge encryption to file transfers, and is honest about the limits of what it can and cannot guarantee: start a free 7-day trial, no card required.

Michael C.

Founder & Principal Engineer, Selina Labs

Michael builds Selina, a privacy-first AI that remembers you across conversations. He ships security-sensitive AI in production — real attacks, real fixes, measured in minutes and dollars — and writes about privacy, security, and LLMs from that seat. Top Rated Plus and expert-verified on Upwork.

• www.upwork.com
• x.com
• medium.com
• selinalabs.io