Where BrainNet Still Fails
On our own vault, every network idea tied plain search at 23 of 29 and 17 of 30. Bridges, usage synapses, learned weights, Hebbian plasticity and a graph inside the model were all killed against a kill line set in advance.
A brain-ask question on the Wikipedia network fell from 5.3 s to 0.69 s because most of the time was loading files, not spreading activation.
On the Wikipedia network, one question used to take 5.3 seconds, and only 0.66 of them was thinking. Most of the rest was loading files. Now the whole question takes 0.69 seconds. You type a question into brain-ask, and about half a second later you get five notes. This article follows the question through that half second, then shows the two smaller tools we use to walk the network by hand.
Two searches run side by side. One is classic keyword search (BM25), which rewards notes that share your words. The other is vector search: the question becomes an embedding, and the closest matching notes win. The vector search uses compressed 2-bit codes for speed, then rechecks the best candidates with exact numbers.
The top 20 results are the seeds. Each seed gets a starting signal of 1 divided by (1 + its rank), so the best seed gets 1, the second 0.5, the third 0.33, and so on.
Now the network does its job. The signal flows from each lit neuron to its neighbours. The method is called personalized PageRank. It is the idea behind Google's early ranking, with one change: the walk always restarts at your seeds, so it stays near the question.
In plain words, each round does two things. Half of every neuron's signal comes from the seeds (the restart). The other half comes from its neighbours, split evenly among them. We run 10 rounds. The final score is the seed signal plus 2 times the spread signal. Those three numbers (0.5, 10, 2) were picked on one small test and then frozen. We did not tune them on the final score set.
A toy example, not from our data. Three notes in a line: A links to B, B links to C. You only matched A.
| round | A | B | C |
|---|---|---|---|
| 0 | 1 | 0 | 0 |
| 1 | 0.5 | 0.5 | 0 |
| 2 | 0.625 | 0.25 | 0.125 |
Note C never matched your words. After two rounds it holds a small signal because it sits two links from a match. That is the whole promise of the network. Whether it is a good promise is the topic of article 5.
The spread gives a re-scored list. A reranking model then reads the question directly alongside each candidate note and scores how well they match. Ours is a small model, ms-marco-MiniLM-L-6-v2, run in Mojo. It picks the final five. In article 4 you will see that this step, not the spread, earned most of the accuracy gain.
On 2026-11-06 we asked: "why did spreading activation not help the Brain vault". The top hit was the 2026-10-21 BrainNet verdict note, its section on scale and gravity tests. Under it came the BrainNet skill file and the PageRank note. That is the right neighbourhood.
We also ran the same question against the wrong store, --net results, which covers math paper results. It returned lemmas that contain the word "spread" ("spread-out onset", "barycentric refinement"). Same question, wrong layer, useless answer. The store flag matters, because each layer answers in its own language.
Asking the Wikipedia network (1,181,724 articles at the time) used to take 5.3 seconds. Our first guess was that it fired every neuron at once. That guess was true: ten rounds from 20 seeds lit 1,175,258 of 1,181,724 neurons, 99.5 percent of the network.
But firing was only 0.66 s of the cost. The other 3.9 s was loading: each question copied a 540 MB link file and a 1.8 GB vector file into memory. Two fixes, which together brought one question to 0.69 s end to end:
The cut-off of 0.0001 passed, 0.001 did not. We kept the one that passed.
Sometimes you do not have a question. You have a note, and you want to know what lives next to it. brain-spread takes one or more notes, runs the same walk from them, and prints the most activated neighbours. It takes about 15 ms. Agents use it in a loop: spread, read what mattered, spread again.
We ran it from the PageRank note. The output:
| activation | neighbour |
|---|---|
| 0.00407 | the graft-net skill |
| 0.00245 | the bridge verdict note |
| 0.00216 | the latency-and-eps note |
| 0.00213 | the first BrainNet verdict |
| 0.00207 | the usage-synapses note |
Numbers are tiny because one note spreads thin over many neighbours. Hub pages link to hundreds of notes and would flood every result. The walk damps their signal by dividing by the square root of their link count. A flag turns that off.
The third tool is the plainest one, brain-recall. It does no maths. It searches every note for your words and prints the matching lines. On 2026-11-06 the query "spreading activation" matched 568 notes, 28 of which contain both words. It is blunt, and that is the point: exact words, no model, no surprises. Our rule is to run it before forming any theory.
| tool | what it does | when |
|---|---|---|
brain-ask |
seeds by meaning, spread, cross-encoder, five hits | you have a question |
brain-spread |
spread from chosen notes | you have a note |
brain-recall |
exact word search with context | you want what we already concluded |
Next: what this machinery fixed in practice.
Reacties
Nog geen reacties.
Inloggen om te reageren.