DARPA built a lifelog in 2003. We ended up with it anyway.
In May 2003, the Defense Advanced Research Projects Agency posted a solicitation with the reference number BAA 03-30. It was addressed to anyone willing to bid on a research contract. The agency wanted to fund a new system, and the system had a name. They called it LifeLog.
The pitch, taken verbatim from the solicitation, was that LifeLog would be “an ontology-based (sub)system that captures, stores, and makes accessible the flow of one person’s experience in and interactions with the world.” The goal was “to be able to trace the ‘threads’ of an individual’s life in terms of events, states, and relationships.” What that meant in practice was spelled out in the same document. LifeLog would collect credit card transactions, web browsing, phone calls, email, physical location via GPS, biomedical readings from wearable sensors, everything the subject read or watched or listened to, instant messages, and postal mail. The pattern-recognition layer was supposed to infer habits, routines, and relationships on top of the raw stream.
It was, in other words, a plan to log a human life.
Nine months later, it was dead. On February 4, 2004, DARPA cancelled LifeLog. The official reason given to reporters was a “change in priorities.” The real reason, as anyone following the news could tell, was that the political climate had turned. Privacy advocates had spent the preceding year in a prolonged fight over a sister DARPA project called Total Information Awareness, and Congress had defunded TIA two months earlier, in December 2003. When LifeLog came up for public attention, there was no appetite left for another program that sounded like total surveillance wearing an academic name.
LifeLog did not die because the technology was wrong. It died because the year was wrong.
What DARPA actually wanted
The thing to understand about LifeLog is that it was not primarily a surveillance program, although in 2003 that distinction was impossible to communicate. It was a research program about memory.
The interesting phrase in the solicitation is “the flow of one person’s experience.” The target user was the subject, not an analyst. LifeLog was supposed to produce something the person being recorded could query: a retrievable, structured version of their own past. The military-adjacent use cases were training aids and personal assistants. A soldier returning from deployment could reconstruct a chain of events. An analyst could remember where a conversation came from. A commander could review a decision by replaying the day it was made.
Those use cases were real, and some of them would later be built by other people for other reasons. But the framing in 2003 made them hard to defend. DARPA funded it. The data would live in a government-adjacent research system. Nothing about the delivery channel, the funding source, or the cultural context made anyone feel good about the idea that someone, somewhere, would hold an ontology-based record of their purchases and their phone calls and their sensor readings. The program could not survive a front-page Wired article, and it did not.
The thing the critics got right
The privacy objection to LifeLog in 2003 was correct. A centralized, government-funded, opaque system that swallowed every data type in a person’s life was not something that could be shipped responsibly in that form, in that decade, under those auspices. The critics were not being paranoid; they were reading the proposal and reporting what it said.
What the critics did not quite say, because it was not yet obvious, was that the objection was not to the idea of a lifelog. It was to who was going to hold it. Put the same data set in a government research archive, and it is surveillance. Put it in a consumer product owned by an ad network, and it is a feature. Put it in a tool that belongs to the person being recorded, and it is a diary.
This distinction did not exist in the public vocabulary in 2003. There was barely any consumer software that captured anything continuous. Gmail was a month from launching. The iPhone was four years away. Facebook was still a dorm room project at Harvard. The only institutions that could plausibly operate a system like LifeLog were governments and a handful of research labs, and the handful of research labs were DARPA-adjacent. In that world, “continuous capture of a human life” could only mean one thing, and the thing it meant was not good.
The twenty years that followed
What happened next is the part nobody planned.
Every data stream that LifeLog proposed to capture now exists. Not because anyone built LifeLog. Because we assembled the equivalent piece by piece, for unrelated reasons, and most of us opted in voluntarily.
Credit card transactions: your bank app has them, and so does every loyalty program and expense tool you use. Web browsing: Chrome knows, Firefox knows, your password manager knows, and the sites themselves know via cookies and analytics. Phone calls and email: in the cloud, searchable, dated, threaded. Physical location: your phone logs it to within a few meters, and Google Timeline will draw it on a map if you ever ask. Biomedical readings from wearable sensors: the smartwatch on your wrist sends heart rate, sleep stages, step count, and blood oxygen to the manufacturer’s servers every few minutes. Everything you read or watched or listened to: Netflix history, Spotify history, Kindle sync, Trakt if you installed it, YouTube’s watch page, the reading-time counter on Apple Books. Instant messages: Slack, Discord, iMessage, WhatsApp, Bluesky DMs. Even postal mail, for people who use Informed Delivery from USPS, is emailed to you as a scan the day it arrives.
Every bullet point in BAA 03-30 is now a tab in someone’s phone.
The data DARPA could not collect in 2003 because the infrastructure did not exist is now produced by default by the infrastructure that does. You do not have to log anything. Logging is the ambient condition.
What DARPA got wrong about the hard part
Read the original LifeLog solicitation today and one thing stands out. DARPA budgeted for capture. They assumed retrieval would follow. They were going to invent sensors, collect streams, and build an ontology. The premise was that if you captured enough, the system would become useful.
That premise was wrong, and it would have been wrong even without the privacy backlash. The lifelogging movement of the late 2000s, which tried to build consumer versions of the same idea, hit the same wall. Gordon Bell’s MyLifeBits at Microsoft Research accumulated terabytes of photos, documents, and audio. The research team itself admitted that very little of it was ever retrieved. The Narrative Clip and Google Glass got further along the capture axis than DARPA would have, and also failed, because storing an undifferentiated stream of photos is not the same as remembering a day.
Capture is the easy problem. Return is the hard one.
For a recorded life to be useful to the person living it, something has to shape the raw stream into a form a human can re-encounter. Not a chart. Not a database. Not a list of events. A paragraph, an image, a moment that comes back. For forty years this was the bottleneck. The sensors were cheap, the storage was cheap, and the composition step was unsolved.
The composition step was unsolved until it was not. Language models changed the economics of turning a day’s fragments into something readable. Not by inventing new data, but by reading what is already there and writing it out as prose.
What shipped in the end
The lifelog that actually exists in 2026 does not look like what DARPA drew up. It is distributed across a dozen services. It is captured as a byproduct of using tools you picked for other reasons. It is retrievable, in its raw form, from APIs those services expose. And it is, crucially, the user’s data, not a central archive.
Which means the question LifeLog raised and could not answer, who holds the record of your life, has been answered differently than DARPA imagined. Right now, the record of your life is held by the individual companies whose tools you use. Each one sees a slice. None of them sees the whole. And the composition step, the thing that turns the pieces into a day you can actually read, has mostly not been built at all.
That is the interesting place we have ended up. The raw material of a lifelog exists, scattered and siloed. The compositional technology exists, commodified and cheap. The only missing pieces are the pipes between them, and the answer to the question DARPA could not answer well in 2003: who holds the composed version, and on whose terms.
The version worth building
The version worth building is the one DARPA could not. A lifelog that belongs to the person being recorded. That never aggregates across users. That does the assembly step on data the user has chosen to route through it, and returns the output to nobody but them. That treats “the flow of one person’s experience” not as a surveillance target but as something the person is entitled to have access to.
In one sense this is not a radical idea. You already own your calendar, your messages, and your commit history. You can export them. The apps that produce them are not built to hand you a narrative, but they are built to hand you your data.
What has been missing is the layer above, the service that takes the data and composes it. In 2003 that service would have had to be a DARPA contractor. In 2026 it can be a small product with a narrow remit: take what is yours, write it into a diary entry, and put the entry in your account, not a shared store.
That is what deariary does. It connects to the tools you already use, reads the day’s activity, and generates a diary entry you can re-read whenever you want. The privacy-first architecture means the raw data and the composed entries live in your account alone. There is no central lifelog. There is no analyst. There is the day that just happened, written down, for you.
What we learned
Looking back at LifeLog, the lesson is not that DARPA was wrong about the idea. The idea, stripped of the funding source and the sensor fetishism and the language of ontology, was that a person could one day have a readable record of their own life, composed from the traces their life leaves behind. Twenty-three years later that sentence sounds unobjectionable, even mundane. In 2003, it could not survive the political climate and it could not survive the technical premise.
Both of those constraints have eased. The technical premise has flipped: the data exists, and the composition step is a solved problem. The political premise has shifted too: a record your tools assemble for you, stored in your own account, is not the same artifact as a government dossier, and the ownership question now has a clear answer.
The lifelog DARPA proposed could not exist in a form anyone wanted. The lifelog that exists now is quieter, smaller, more honest about whose data it is, and delivered one ordinary day at a time. It is the version worth keeping. It is, finally, an automatic memory that does not belong to DARPA, or to a platform, or to an analyst somewhere. It belongs to you.
