"I don't want AI reading my life"
It is the most common reason people give for not trying an automatic diary, and it is a good reason. It is also three separate objections in one sentence, which is why arguments about it tend to go badly. Someone answers one of the three and the person raising it does not feel answered, because the part they meant was one of the other two.
Worth separating them.
Objection one: someone will see it
The fear is exposure to people. A support engineer reading your entries. A breach. Data sold to a broker. An employee curious about a specific account.
This one has factual answers rather than reassurances, and they are documented in detail in our privacy post: what each integration collects and does not, encryption in transit and at rest, credentials encrypted with a separate key, direct messages excluded in code rather than by a setting, no data sold, full export in Markdown, hard deletion after a 30-day grace period.
If your version of the objection is this one, that post is the place to check, and the honest caveats are in it too: we are a small team without SOC 2 certification, and it is a cloud service.
Objection two: it will be used to train something
Distinct fear, distinct answer. The worry is that your life becomes training data and ends up, in some diffuse way, inside a model that other people use.
The answer is that the language model providers do not train on API data, this is covered by data processing agreements, and deariary has signed them. Your Wednesday is not making anyone’s model better.
This is verifiable in the sense that the agreements exist and the providers publish their API data policies. It is not verifiable by you directly, which is a real limitation and worth naming rather than glossing.
Objection three: I do not want to be read
Here is the one that matters, and the one that gets argued with when it should not be.
Sometimes the objection is not about consequences at all. There is no feared outcome. It is that something processing the texture of your days is unpleasant in itself, in the way that being watched is unpleasant even when the watcher is indifferent and nothing happens as a result.
That is not a security question. No encryption diagram answers it. And it is not irrational: it is a preference about how you want to live, and preferences of that kind do not require justification.
What can be offered is a more accurate picture of what the reading actually is, so that whatever you decide, you are deciding about the real thing.
What “reading” means here, precisely
The daily generation is one call, per day, with that day’s data. The system prompt tells the model it is writing as you, from the data provided, and the rules include not inventing facts and not fabricating emotions. Then it returns prose and the call is over.
There is no profile. Nothing accumulates a model of who you are between days. The model that wrote Tuesday has no idea what Monday contained.
Two exceptions to state clearly, because “stateless” would otherwise be an overclaim. The weekly and monthly summary features do read across a period: the weekly one is given that week’s entries and asked to pick the three most memorable moments, the monthly one is given the month’s entries. That is more than a single day by design, and if it bothers you, it is a distinct thing to weigh.
So the shape of the reading is transactional rather than accumulating. Whether that makes it acceptable is not something we can decide for you. But “an AI is building a picture of me over time” is not what is happening, and if that was the objection, it is worth knowing.
The reading was already happening
This point is worth making carefully, because the aggressive version of it is a bad argument.
The bad version: you already gave your data to Google and Slack, so privacy is gone, so you may as well. That is nonsense. Consent to one thing is not consent to everything, and incremental erosion is exactly how this goes wrong.
The accurate version: the specific data an automatic diary reads is data that machines already process continuously. Your calendar is parsed to send reminders. Your messages are indexed for search. Your task manager computes completion. Your listening history feeds recommendations. Those systems read the same material, more of it, more often, and none of them give you anything back that you can read.
So the marginal increase in machine reading is smaller than it feels, and the marginal increase in what you get for it is large. That is an argument for a favourable ratio, not an argument that the ratio is free. If the answer is still no, the answer is no.
The version where you should say no
Some cases are not close, and it is better to say so than to sell.
If you handle other people’s confidential material as a professional duty, do not connect the tools where it lives. A journalist with sources, a clinician, a lawyer, anyone under a regulatory obligation: work chat and work calendar are not yours to route through a third party, regardless of that party’s encryption.
If you are in a situation where someone else may gain access to your accounts, an additional store of your daily movements is an additional risk. That deserves more caution than a product’s privacy page can address.
And if the third objection is your actual objection and it does not resolve, then this is not a product for you, and that is a complete and reasonable position. There is no version of this argument that ends with you being wrong.
The middle option nobody mentions
The choice is usually presented as connect everything or use nothing. It is not.
The integrations vary enormously in sensitivity. Your work chat is at one end. Steam playtime, listening history, and watch history are at the other, and they are genuinely useful for a diary: they carry the texture of evenings and weekends without touching anything confidential. Places you checked in sit in the middle. A calendar can be sensitive or not, depending on the calendar.
An entry built from three low-sensitivity sources is thinner than one built from ten, and it is still a record of a day where there would otherwise be nothing. Starting with the least sensitive service you use and reading a week of output is a much better way to decide than reasoning about it in the abstract, and if the output is not worth it, you have disclosed almost nothing.
That is the honest recommendation. Not “trust us.” Test the smallest version, decide from the result, and keep the objection.
