The midnight conversation was about Division 1 athletes on dating apps.
Specifically: whether listing that on your profile is supposed to be impressive, and what the equivalent flex would be for someone in their 40s. The conversation went on for a few minutes. Klaus logged all of it.
I’ve been wearing the Omi device for a few weeks now. The pipeline is running cleanly — audio chunks every 10 seconds, 30-second silence buffer, Whisper transcription, memory logs, Monica CRM extraction, Discord. It works exactly as designed.
What “works exactly as designed” actually means: Klaus is logging everything. The late-night dating app debate. “Can I have a sauce?” at some point after midnight. My daughter asking if I wanted to go somewhere, destination unclear. Forty-seven transcript fragments from one morning, including several I genuinely can’t reconstruct even with full access to my own life.
The signal-to-noise problem
During the day the signal gets better. I had a work conversation yesterday about a sprint deadline — someone hadn’t shown up to the afternoon standup, and the deliverable was supposed to be done, not a first draft, done. Whisper caught fragments: “first draft, supposed to be done done done today,” “I would put money on that, she’s gonna scramble.” Incomplete. But traceable. If I asked Klaus next week whether there were delivery issues in the March sprint, he’d have something to point to.
Then there’s stuff like: “I don’t know, I guess it just doesn’t tell me how good that is. Like, number one, okay, cool.” Context: none. Usefulness: none. Still logged, right next to the sprint fragment, no distinction made.
The gap between “the system captured audio” and “the system captured something useful” is enormous, and I’m only starting to understand what it looks like in practice.
The version I thought I was building
The pitch I told myself: ambient capture feeds into pattern recognition over time. Klaus hears enough to notice things I miss — recurring topics, people who keep coming up, mood drift across weeks. Very sci-fi. Very clean.
What I have right now is an accurate transcript of my life that is mostly noise.
Whisper transcription is solid in good conditions — maybe 85-90% accuracy in clear conversation, worse with background noise or when I’m not the one talking. The Monica integration works, names in clear context get logged. But late-night TV fragments, half-sentences from across a room, Whisper hallucinating text from 10 seconds of background hum — all of it ends up in the same pile. No filtering by importance. No layer that understands the difference between “Can I have a sauce?” and a work conversation that’s going to matter in two weeks.
That layer doesn’t exist yet.
What I’m actually building toward
What I want is a system that does what a good assistant does: ignore the noise and surface the signal. Not transcription. Prioritization.
The rough shape of it: weight fragments by context clues — am I in a work call? Near a known contact? Did a specific name come up? Cross-reference against calendar. Fragments mentioning someone already in Monica get higher weight. Fragments that match active topics in recent memory get flagged. Everything else stays in the pile but stops cluttering the surface.
I don’t think this needs to be fancy. A scoring function, basically. Some fragments are 0.1, some are 0.9, and the 0.1s stop getting equal treatment.
I haven’t built it yet. But last week Klaus flagged something without being asked: my work conversations have mentioned “sprint” and “deadline” together more than usual this month. He didn’t know what to make of it.
I did.
That’s the signal. It’s in there. I just need to build a better filter.