The Monthly Letter: What an AI Can Reflect Back That You Can't See
The Monthly Letter: What an AI Can Reflect Back That You Can't See
Bottom line: you are a bad witness to your own last thirty days. Not because you're careless — because human memory of emotion is a reconstruction, and it distorts in known, repeatable directions. A monthly letter built from what you actually wrote at the time is worth having for exactly that reason, and for no other reason worth paying attention to.
I build a journaling app that writes one of these letters. So read this as a biased source. I've put the research and the failure modes in, including the ones that apply to mine.
Your memory of last month is a summary, not a recording
Four findings, all boring, all well established.
Peak-end. You don't remember how an experience felt on average. You remember its most intense moment and how it ended, and you file that as the whole thing. Kahneman's work on remembered versus experienced utility is the canonical version. A rough month with a good last week gets remembered as a decent month. A fine month that ended badly does not.
Fading affect bias. Negative feeling attached to a memory decays faster than positive feeling. This is generally good for you — it's part of how people stay functional. It also means that when you look back, the bad stretch genuinely feels smaller than it was while you were in it.
Recall is not symmetric. Studies comparing what people report in the moment against what they report later find systematic over-estimation of negative emotion in retrospect, particularly anger — and the effect is strongest in people with a history of depression. So depending on who you are and what you're looking back through, the same month can compress in opposite directions.
You misremember your own predictions. This is the one that gets me. Research on affective forecasting finds that people recall their past predictions as having matched what actually happened. Their forecast error becomes invisible, so they never correct it. Remind people what they actually predicted and they start forecasting better. Don't, and they repeat the same mistake for years.
Put those together and you get an uncomfortable conclusion. Reviewing last month from memory is not reviewing last month. It's reviewing a compressed, peak-weighted, affect-faded story that your brain wrote about last month — and the errors are not random, so they don't cancel out over time. They accumulate in one direction.
What a letter can do about that
The fix has been known for decades and it isn't AI. It's concurrent recording. Write things down near when they happen, then compare the record to your memory. Psychologists use this method precisely because retrospective self-report is unreliable.
A journal is already a concurrent record. Most people just never read theirs. Four hundred entries sitting in an app do nothing, because re-reading four hundred entries is work nobody does on a Sunday.
That gap — you have the record, you'll never read it — is the only honest job description for a monthly letter. Not therapy. Not advice. Not a friend. A readable version of a record you already own but won't sit down with.
Which means the useful question isn't "what does the AI say about me." It's narrower:
What did I write about more than I realised? Frequency is invisible from the inside. You do not know that you mentioned the same person in nineteen of thirty entries. You'd swear it was four.
What faded? The thing that consumed three weeks in June and that you cannot now name. That's fading affect bias with a receipt attached.
What did I say I'd do? Your own words from four weeks ago, before you knew how it turned out. This is the forecasting-error mirror, and it's the single most valuable line in any letter I've read from my own journal.
What changed shape? Not "your mood improved." That's a chart, and charts of your own mood mostly flatter you. Something more like: the thing you wrote about with dread in week one, you wrote about with boredom by week four.
Notice none of that requires the software to be wise. It requires it to count, quote, and stay out of the way. Most of the value in a monthly letter is retrieval, not intelligence.
The failure modes, including mine
A letter that flatters is worse than no letter. Summaries drift toward "you had a full month and showed real resilience." That is horoscope output. If your letter never says anything you'd rather not read, it isn't reading you — it's reading a template.
Insight is not a diagnosis. A journal that notices you wrote about sleep eleven times has noticed that you wrote about sleep eleven times. It hasn't found anything clinical, and any app implying otherwise is doing something it shouldn't. Journaling is a practice, not treatment, and nothing in a monthly letter substitutes for talking to a professional.
A summary is a second copy. This is the part the category skips. To write a letter about your month, something has to read your month. That's a real privacy event, and "we generate monthly insights" tells you nothing about where the reading happened, who could see it, or whether the letter and the entries are protected the same way.
So the checklist for anyone's monthly summary — mine included:
- Where does the summarising happen, and what can a human at that company see while it happens?
- Is the letter stored with the same protection as the entries, or does it quietly become a plaintext digest of your worst month?
- Are your entries used to train anyone's model?
- Does it quote you, or does it characterise you? Quotes are checkable. Characterisations aren't.
Running it on my own app
Uncloud writes a monthly letter. Here's how it answers its own checklist, in plain terms.
The AI work happens on our servers, not on your device — I'm not going to pretend otherwise, and any AI journal that reflects across months and claims full end-to-end encryption is worth a second look. What we do instead is narrower and checkable: your entries are encrypted so completely that even we can't read them, and that isn't a policy promise — administrative read access is blocked in the code and enforced by automated tests that run on every build. If someone breaks it, the build fails. Your entries are never used to train models; that's contractual with our AI providers, not a preference.
The letter is built from your own entries and quotes them back. Everything stays on your device, syncing only if you want it to. It works offline. And it will occasionally tell you something you didn't want to know about August, which is the entire point.
That's the honest pitch. Not that an AI understands you. That you already wrote down more than you remember, and something should read it back to you accurately once a month — including the parts your memory has been quietly editing.
Uncloud is free to start on Android: Uncloud: Private AI Journal on Google Play. Write for a month. Then read what you actually said.