TRACE turns the world into words the instant it happens — never keeping one frame of video or one second of audio — binds those words into living nodes at night, and answers any question about your day with citations, or an honest “I don't know.”
The particles behind this text are observations firing. Everything on this page is live — scroll.
Everything a mind knows arrives through senses. For a machine, the only senses we can cheaply manufacture are eyes and ears — and most of the information worth having flows through them anyway. TRACE copies that shape exactly, then makes one radical amendment: the brain keeps only what the senses said, never what they saw.
You don't remember pixels. You remember that a crow was on the railing — a fact, located in time and place.
Raw pixels and audio live for milliseconds — just long enough to become typed text. Then they're gone. There is nothing to leak, subpoena, or replay. Surveillance isn't forbidden; it's structurally impossible.
This is the napkin, industrialized. Each stage lights up as it would fire on a real moment. Hover any stage; the napkin's own words are quoted underneath.
Video (eyes), audio (ears), plus motion, pose and place — the cheap channels run continuously.
"Video (Eyes), Audio (Ears)"Blurry frames discarded, signal deciphered from noise. Out: coordinates, timestamps, context.
"Extraction Black Box"A small vision-language model gives the gist of what's happening — only when attention, dwell or change earns the cost.
"A macro picture of what's happening ⇒ VLM"Specialized eyes called per task: OCR reads text, ASR hears speech, a tagger names sounds, a detector tracks objects.
"A specialized eye or helper for various tasks, called as needed"The best understanding possible of the world — every line marked by coordinates and time. This is all that survives.
"TEXT — text from various helpers marked by coordinates"Binds coordinates, physicality and temporality together: which hum was the fan, which blue thing was the bag.
"Binder ⇒ binding the coordinates, physicality & temporality together"Raw material is marinated into living nodes that combine with fellow nodes to derive new things. 1+1=4, not 2. Encrypted — only your app deciphers.
"A physical living breathing cell… always running, but better nightly"Assembles just enough context for the local LLM. The LLM does what LLMs do — the value is everything before it.
"Just Enough Context ⟶ LLM does what LLM does"The discarder rule: marinated content that proves useless is dropped — never the raw material. The append-only observation log is forever; everything derived from it is disposable and replayable.
Tuesday morning. A crow lands on the balcony railing while you take the trash out. No video is kept — yet three mornings later, TRACE knows something nobody told it. Step through what actually happens. Nodes that fire together, wire together.
The cheap channels are already running — detector, sound tagger, pose. They cost almost nothing, so they never sleep. The expensive eye (the VLM) is dark, waiting for something worth looking at.
A day of memory as it actually rests in the store. Run a day to stream observations in; run the night shift to watch the binder marinate them — strong co-occurrences harden into green bindings, weak links flash red and are refused, not hardened. Hover any node to read the words it holds. Drag nodes to feel the graph.
Every node carries provenance — the event id, channel and millisecond it came from. A claim without a source is a construction error, not a style violation.
Frames arrive, words leave, pixels die at the wall. The type system is the
law: the only object that may cross the device boundary is TextEgress — typed
text carrying the ids of the observations it came from. Anything else raises
PrivacyViolationError. Watch the membrane work.
Nothing here is vibes. A blind adversarial question battery runs against real captures; fabrication is subtracted, not just un-credited; and every failure is mechanically diagnosed — (A) never captured, or (B) captured but the brain failed to use it. Helpers are only built when a missed question demands one.
Correct minus made-up, over total. A confident wrong answer costs what a right one earns. An honest "I don't know" is a win.
Of the questions an oracle with the raw frames can answer, the share the retained text cannot. The direct price of throwing the pixels away — driven toward zero.
Hallucination approaches zero across every phase. Current honest baseline: 8 of 9 misses were the brain, not the sensors — so that's where the work goes.