Most AI journaling apps have no memory at all. You write an entry, a model responds to that entry, and the exchange ends there. Tomorrow it starts over. The app has read a great deal of your writing and remembers none of it.
That is a strange design for a product whose entire premise is that writing accumulates into something. An AI journal with long-term memory is the attempt to fix it — and the interesting question is what “memory” is doing in that phrase.
Three things “long-term memory” can mean
A longer context window. The model sees more of your recent entries at once. This is the most common implementation and the weakest: it extends how far back the app can look in a single sitting, but nothing is retained between sittings.
Stored summaries. The app keeps a compressed account of previous periods and refers back to it. Cheaper, more durable, and it survives across sessions — but you are being read a summary of yourself.
Semantic connection across time. The app can find an entry from two years ago that resembles this week, not because the words match but because the meaning does. This is the version that produces the result people actually want: you wrote about this before.
The third is the hard one, and it is where the phrase earns its keep.
What to ask before you commit years to one
Does it connect across years, or across days? An app that can only reach back a few weeks does not have long-term memory in any useful sense.
Is the memory yours or the model’s? There is a difference between an app that resurfaces your own entry and one that paraphrases it back at you. The first is retrieval; the second is a summary you cannot check.
What is sent, and to whom? Semantic connection over years requires processing your writing somewhere. Every honest answer to this question names a provider.
Is it on by default? For a journal, the answer should be no.
Does it work if you turn the AI off? The strongest products keep something useful when the model is not involved.
How Deeditt handles it
The connection is made on summaries, not raw entries
When you turn Insights on, your journal is read back to you at four distances: a week, a month, a quarter and a year. Each comes with a short written account of what you lived, in your own terms, and with the people, projects and themes that run through the period and where they stand.
Alongside them come echoes: your own earlier words, returned when they are relevant again. What you wrote on this date a year ago. A line from months back that resonates with this week. Two entries about the same thing, far apart, that pull in opposite directions. That is what long-term memory looks like in practice: not a search box, but your past self showing up at the right moment.
What leaves your account to make this possible, and what never does, is set out plainly in what happens when you turn on Insights.
What is actually sent
When the analysis runs, the payload is the current and previous week’s entries — the dates and the lines you wrote, and nothing else. No user id, no username, no email, no account identifier. OpenAI produces the weekly reflection and the recurring patterns; Cohere produces the embeddings that connect distant weeks.
Two claims often get blurred together here, and they should not be. That your writing is not used to train external models is true, and it is guaranteed by the contract with those providers — a commercial commitment. That your entries are encrypted at rest is also true, and it is guaranteed by mathematics. Both are real. They are different kinds of real, and any product that presents the first as though it were the second is misleading you.
It is off until you turn it on
The setting is off for every account from the moment it is created. It is not a preference we default to on and let you find later. The weekly pipeline filters on that flag before anything is decrypted, so accounts that have not opted in are excluded before any content is opened at all.
If you never enable it, no entry of yours is ever decrypted for analysis. Turning it off stops future analysis.
And the part that works with no model at all
The feeling, sleep and energy you record are counted from what you tapped. No model reads them, nothing is sent anywhere. So that part keeps working whether or not you ever switch the written Insights on — which we think is how it should be.
What it costs
Insights, at all four horizons, are part of Pro, $4.99/month or $49.99/year. Writing in Memories is free on every plan. The boundary follows the cost of running things: writing an entry is cheap, and the analysis built on top of it is not.
The honest limits
Removing identifiers does not make a journal anonymous. We are not going to call it anonymised. An entry that names a manager, a city and a medical procedure points at one person whether or not an account id is attached. Journal text is made of exactly the details that identify a life. That is the real reason the setting is off by default, and it is the answer we will give whenever the default is questioned.
Deeditt is not end-to-end encrypted. Entries are encrypted at rest and in transit and are private from other members, and the encrypted content, the keys and the passphrase that unlocks them are each stored in a different system. But keys are generated on our servers. If you need a provider that is cryptographically unable to read your journal, an end-to-end app is the right choice — and it will not be able to offer this feature, because nothing can analyse what it cannot decrypt. That trade-off is real and it is yours to make.
It reads weeks, not moments. The reflection runs on completed weeks. It is not a conversational assistant and it will not respond to the entry you just wrote.
What it is actually good for
The value is not that a model says something clever about your week. It is that after a year or two, something you wrote in a completely different context comes back at the moment it turns out to be relevant — the version of you who already solved this, or already failed at it, handing over notes.
That only works if the writing is there in the first place. The memory is the feature; the entries are the thing that makes it possible.
