"Hey ChatGPT, remember that I'm pivoting the agency toward AI consulting next quarter."
"Remember that my mom's chemo starts Thursday."
"Remember that I'm not sure I want to stay in this relationship."
People talk to AI assistants the way they used to talk to journals. The difference is that the journal didn't have a business development team.
The quiet shift from tool to confidant
A year ago, ChatGPT was a search engine you could argue with. Now it has memory. It knows your job, your preferences, your recurring projects, the names of your kids. Google's AI products are racing the same direction. The framing is always the same: this is for you, to make the assistant more helpful, to save you repeating yourself.
Some of that is true. The part they don't emphasize is that "helpful" and "monetizable" are not different datasets. The exact record that makes the assistant better at you is the record that makes you better to predict, segment, and sell to.
What "product improvement" actually covers
Every consumer AI product has a clause about how your data is used to improve the product. That phrase is doing a lot of work. "Product improvement" can mean making the model smarter. It can also mean making the funnel sharper. It can mean training a future advertising product. It can mean sharing aggregated, "anonymized" insights with partners. It can mean almost anything, because the policy is written by the company, revised by the company, and enforced by the company.
When the dataset in question is "everything you've told the assistant about your life," the vagueness is not a technicality. It's the whole game.
The ad-targeting pipeline you didn't opt into
You didn't sign up for an ad product. You signed up for a chatbot. But the memory feature is, structurally, the richest first-party data collection mechanism invented since the newsfeed. It's not clicks. It's not dwell time. It's stated intent, in your own words, dated and contextualized.
If that sounds like the input to an ad-targeting system, that's because it is. Whether any given lab uses it that way today is almost beside the point. The asset is being accumulated. The optionality is the value. You are building it for them, one "remember this" at a time.
The asymmetry, again
You can't see what they've inferred about you. You can't audit the profile. You can't export the underlying model of your personality and take it to a competitor. You can delete individual memories, sometimes, if you remember where the setting lives. You cannot meaningfully consent to a system whose downstream uses are unspecified and changeable.
This is not a privacy policy problem. It's a structural one. The memory lives on their infrastructure, under their control, optimized for their goals. Your goals and their goals overlap just enough to keep you engaged, and diverge everywhere it matters.
The consent problem nobody can solve
The standard answer is "you can turn it off." But turning off memory after you've been using it for months is not the same as never having turned it on. The model has already absorbed your patterns. The profile has already been built. The "off" switch stops future collection; it does not undo what was already collected, and it does not give you a copy of what was inferred.
Real consent would require three things: knowing what's being collected, knowing how it's used, and being able to take it with you when you leave. Consumer AI memory fails all three. The first is buried in settings. The second is buried in policy. The third isn't offered at all. You can consent to the checkbox. You can't consent to the system.
A different architecture
Eigen Mesh exists because we think the architecture is the point. If the memory lives with the company, you lose. If the memory lives with you, the incentives line up.
Eigen Mesh is an AI-native second brain. You capture what you think and plan — in your words, when you mean it — and it organizes that context and offers it to whatever AI tool you happen to be using. The assistants come to your memory. Your memory doesn't get uploaded to theirs.
We don't train on your data. Not because we're nice — because the architecture doesn't require it, and the second we did, we'd be the thing we built this to escape. Apache 2.0, self-hostable, servers in the EU. You can run it yourself and the question of trust becomes moot.
The small, honest version
You don't have to believe AI companies are evil to see the problem. You just have to believe they're companies. They will do what companies do: optimize for the metrics they're accountable to. If those metrics include "engagement with memory features" and "data available for future products," your private context will be used accordingly, no malice required.
The fix isn't a better policy. It's a different place for the memory to live.
If you want your AI to remember things about you, that's reasonable. Just make sure it's remembering them in a place you hold the keys to.
Eigen Mesh is open source — github.com/hikaru90/eigen. Get early access to the hosted version, 100 free credits to start, then €3.99/month. Your memory, not theirs.