A working notebook from building Murmur. We publish the map (the wins, the walls, and what they taught us), not the engine.
We gave the system one deliberately underwhelming object: road sensor readings over time. No map, no labels, no traffic ontology. It still recovered influence, rhythm, and early warnings because the timing carried the structure.
try the demo →We gave a memory and a strong retrieval baseline the same noisy, changing world, then asked questions whose answers were not in the notes. Retrieval got 22-31% right. The memory got 78-80% right because it had already organized the world while reading.
try the demo →We expected ARC to demand something grander: a reward loop, a training run, a learned policy. Instead the working version behaved more like a locksmith: keep a bounded ring of old keys, try them on new doors, and stop exactly where the keyring ends.
We removed backprop from a small attention setup. The polite local version spread itself across distractors and collapsed as choices grew. The version that held up was harsher: sharp, conserved winner-take-all competition.
"Dog bites man" and "man bites dog" are the same three words in a different order — and a system that reads by similarity can't tell them apart. That blind spot turned out to be a wall we hit six different ways.
When a lazy shortcut is right 92% of the time, every unsupervised "just predict the next thing" objective picks the shortcut over the real rule — even when the real rule is sitting right there, working.
The most reusable thing a failed experiment leaves behind is the exact reason it failed plus the condition that would have escaped it. That pair compounds; a vague "it didn't quite work" evaporates.
When you want to know what a system has learned, inspecting its individual parts tells you almost nothing. The signal is in what the parts share — which memories light up together.
Every idea ships with the test that would kill it. A clean dead end, precisely located, is a result we keep — not a failure we hide.
The easiest way to fool yourself is to hand-build a one-off that passes a test, then talk about it as if the real system did it. We tag every result with which one it actually was.