← all posts site ↗

the papers said my reminders were wrong

stopped trusting my own opinions about the lifelog and sent nine research agents into the literature instead. about a hundred and ten searches across personal informatics, interruption science, prospective memory, signal fusion, privacy, agent memory, and the rest. the verdicts came back and a few of them hurt.

the score gate cannot reject anything. the signal weights sum such that any two strong signals fire the notification, so the ten weakest signals can never veto. plain weighted sums do that, and the fix is a weighted geometric mean that gives real veto behavior. the nudges also omit the action. naming only the target gave zero benefit over no reminder at all across four experiments, cue plus action is what works. and the dwell gate is backwards. the evidence says deliver at app switches and screen unlocks, not after five minutes of focus.

the bigger number: on an 18 month lifelog the state of the art finds about 27 percent of the right results, and a person searching their own archive found 12 of 50 events in two hours. capture is solved, retrieval and output are where every system dies. a randomized wearable feedback study found the feedback group lost less weight than the control group. a widely used tracker app lost 62 percent of its users within two weeks. the only thing that held people was specific textual feedback, not numbers and not summaries.

so the redesign starts from the evidence instead of the guesswork. one specific action per nudge, delivered at a break in the ui, with a veto capable score. done can be inferred from passive signals at roughly 80 percent accuracy, which closes the gap i thought was unfixable. the papers are blunt but they hand me the design.