Remembering What You Care About
In September 2025, OpenAI shipped almost exactly the product I keep describing when I talk about living knowledge. ChatGPT Pulse researched your interests overnight — drawing on your chat history, memory, and connected calendar and email — and met you each morning with five to ten visual cards about the things you care about. Sam Altman called it his favorite addition to the product. Nine months later it was gone: in June 2026, as reported, OpenAI sunset Pulse and pointed users at upgraded scheduled tasks instead — briefings you configure yourself, on a cadence you choose. The most-resourced attempt yet at proactive AI memory retreated to a cron job.
I have to take that seriously, because the idea Pulse embodied is one I badly want to exist. The dominant loop today is: I remember a question, I ask, the AI investigates, the question dies with the chat. I am the scheduler of my own curiosity, and I am bad at the job — the standing questions I actually care about get re-asked at random intervals, or never. The flip I want is: I express a durable interest once; the system remembers it; the world changes; the system notices that the change matters; it comes back to me. The important form of memory stops being “the AI remembers me” and becomes the AI remembers what I care about while I am not thinking about it.
Pulse’s sunset is the strongest available evidence against that flip. So this essay is built around the counterevidence, not in spite of it.
The graveyard has a pattern #
Proactive assistance has been shipped, repeatedly, for thirty years, and it keeps dying the same way. The Remembrance Agent (1996) continuously watched what you typed in Emacs and proactively surfaced relevant old notes and mail, deliberately nonintrusively. Microsoft’s Lumière project built Bayesian machinery for inferring a user’s goals and the expected utility of stepping in — and then, by Eric Horvitz’s own account, the shipped Office Assistant dropped the utility calculus and fired on a simple rule system, constantly. That is how you get Clippy. Google Now (2012) put anticipatory cards on millions of phones and was absorbed into Assistant and de-emphasized. Now Pulse.
Meanwhile the boring sibling keeps winning: Google Alerts, RSS, scheduled tasks, scheduled routines. Proactive push keeps dying; user-scheduled pull keeps shipping. That is the market verdict, three decades running, and any essay advocating proactive memory owes it an explanation.
The measurements agree with the graveyard. The best published model for deciding when to offer assistance — fine-tuned against ProactiveBench’s 6,790 human-labeled events — reaches an F1 of 66.47%. That is state of the art, and it is wrong roughly a third of the time. A proactive assistant that is wrong a third of the time is not an assistant; it is a nag. (ProAgentBench, a February 2026 preprint I would treat as a first observation rather than an established result, at least decomposes the problem correctly — timing separately from content — and finds that long-term historical context specifically improves the timing decision. The skill that is missing is exactly the skill memory should feed.)
The principle was solved in 1999; the folklore wasn’t #
Here is what makes the graveyard maddening rather than merely discouraging: the governing principle has been known since before most of the bodies were buried. Horvitz’s mixed-initiative principles (CHI 1999) and his attention-sensitive alerting work (UAI 1999) state it as decision theory:
In words: speak only when what you are bringing is expected to be worth more than the attention it will destroy, given what the person is doing right now — and both quantities must be estimated under uncertainty, per notification, not assumed. Clippy is what you get when you delete the right-hand side. As far as I can tell, most current proactive-agent work still ignores this inequality; the papers optimize what to say and barely model the price of saying it.
The right-hand side is real and measurable. In a field study of alert-driven task switching, Iqbal and Horvitz found users spent about ten minutes on the switch itself and another ten to fifteen wandering through other windows before resuming the suspended task.
But the attention literature also carries a warning about how easily this discourse runs on folklore. The famous statistic — “it takes 23 minutes and 15 seconds to recover from an interruption” — appears in no peer-reviewed paper. It traces to a 2006 Gallup Business Journal interview with Gloria Mark, describing time-to-resumption in field data, and recent investigations found that the papers it is routinely attributed to do not contain it — while Mark’s own most-cited study found interrupted work is actually completed faster, at the price of significantly more stress. I take two lessons. Interruption cost is real but subtler than the slogan. And the burden of proof runs both ways: nobody has published a controlled measurement showing that proactive AI updates net-conserve attention either. Both sides of this argument are currently running on vibes.
Which is why the thesis statement for this whole design space is Herbert Simon’s, from 1971: a wealth of information creates a poverty of attention — and, from the same essay, an information system should be measured by the attention it conserves, not the information it delivers. A proactive memory that cannot demonstrate net attention conservation has no reason to exist.
Pull with memory is most of the flip #
So was the concept falsified, or was it early? Two readings of Pulse’s death are consistent with the public record. One: users genuinely want to remain the scheduler, and the flip is a designer’s fantasy. Two: Pulse failed on cost and relevance rather than concept — a compute-heavy overnight feed, initially for $200-a-month subscribers, pushing generic cards with a when-to-speak skill that the best benchmarks peg near 66% — the way handwriting recognition failed for decades before it worked. I lean toward the second reading, but I want to be plain that this is a hypothesis; the cohort data that would settle it sits unpublished inside OpenAI.
Notice, though, what the “winning” alternative actually is. A scheduled daily briefing already implements most of the flip: the system remembers the question, keeps the context, and does the work unprompted. Only the cadence is still human-set. The remaining distance between a cron job and a living memory is precisely the Horvitz inequality — replacing the timer with an escalation decision gated on expected value, so the system fires when the evidence warrants rather than when the clock does. That is the architecture I sketched in micro-researchers: standing questions, cheap scheduled observation, and escalation to a human only when accumulated evidence crosses a threshold. Its notification layer should behave like calm technology — living in the periphery of attention and earning, rarely, the move to the center. What has genuinely changed since 1999 is that the escalation predicate can now be language-level cognition — “tell me when something changes my conclusion,” not “tell me when a keyword matches” — an argument I make in the topics-as-processes post, with the refresh mechanics in the living knowledge base post.
One objection I cannot fully answer: interests drift. A system faithfully optimizing yesterday’s declared interests is a filter bubble with a scheduler, and information foraging theory suggests a subtler cost — humans learn the scent of relevant information by foraging for it, and a reader of pre-digested diffs may slowly lose the nose. The watch list needs a watcher; who refreshes the refresher is an open design problem, not a footnote.
What would change my mind: Published retention data showing that users offered a genuinely value-gated proactive system — one that fires on the Horvitz inequality rather than a feed or a timer — still migrate back to self-scheduled pull would convince me the concept, not the execution, is what keeps failing. So would two or three more well-funded proactive products dying after demonstrably modeling interruption cost, or when-to-assist accuracy climbing well past 90% F1 with acceptance rates staying flat.