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.
Agents only start from what we already know if a rule makes them recall before they propose anything new.
Our automated coding assistants start each task with empty context. Their first job is to query the memory network before suggesting a fix. They have a shell and a file system, and the network exists so that they do not start from zero. This article shows how they use it, and what rules keep them honest.
A memory store is worthless if an assistant guesses instead of looking. Before stating any root cause for a problem that could have history, the agent must recall first. The method:
brain-recall with the topic words. It searches the whole vault and prints each matching note with its date, its hit count and the matching lines.The guide tells a real story behind the rule. An agent spent hours re-deriving a graphics-card fault and had the human reboot hardware. The answer, an undervolt of 100 millivolts, was already written in three older notes. The agent read one, missed the other two and guessed before recalling. The gate was built from that incident.
This is where the network helps. Exact search finds what you thought to name. The network finds what sits next to it.
The vault holds memory: what happened, dated and never rewritten. It does not hold how the code works today. That lives in each project's own docs, next to the code. Notes link to the docs and never restate them. Mistakes go into a ledger, one file per domain. State to resume goes on a board.
The rule matters for a network. If a fact lives in three places, three neurons answer and two are stale. With one home, links do the joining.
For an agent, brain-spread is a cheap step. It costs about 15 ms. The loop is: ask a question, read the best note, spread from that note, read what lights up, and repeat. We ran it in article 3 from the PageRank note and it surfaced the bridge, latency, usage and verdict notes beside it, with no keyword in common.
The same abilities are exposed to agents as tools in the local hub: brain_ask (it takes a choice of notes, papers or both), brain_recall and brain_spread. The agent that gets a plain-text question and the agent that gets a note path can use the same network.
Two hooks run on skill neurons. After an agent loads a skill, the first hook spreads from it and suggests three companions. We saw memory-optimization suggest three optimization skills for Rust, C++ and ROCm. The second hook runs on each user prompt, asks the network which skill fits and writes the answer to a shadow log without injecting anything. It is off the critical path on purpose: we have not measured whether its suggestions are right, so it only watches. We have not yet measured whether these prompts pick the best skill.
Our rule: each agent turns the repository it works in into neurons. A background script indexes code by splitting files into 220-word chunks, generating embeddings on the GPU, and adding them to the network. repo-net hook runs on each commit. The merged rebuild behind it runs at most once per 30 minutes, and later changes wait for one trailing run. Then you can ask the network a question about code: brain-ask --net text "is there already a ...".
A real example from 2026-11-06. We asked the text network "is there already a function that rebuilds the merged network". The top hits were two notes from an earlier agent that had changed the refresh step and checked that a rebuild from temporary inputs gave byte-identical files. That was the right place to start reading.
Two helpers sit on top. code-links.sh ties chunks to each other with a call graph and to the notes that name them. code-lessons FILE lists the notes that mention a file, so an agent sees past mistakes before it edits.
The code atlas beside it counts 262,032 own symbols over 358 repos, plus 1,024,026 vendored symbols over 102 more. The reuse order is words, then meaning, then opening the file.
An agent that has never seen the machine needs a front door. net-index --write reads the merged network, each store's edge list and the Wikipedia build log, and writes a single map file in 0.4 s. It lists layers, stores, neuron and synapse counts, and how to walk the graph. We regenerate it rather than edit it, so the numbers in it are true on the date it was made. Many of the numbers in this series come from it or from the notes it points to.
We ported the tools to an older dual-core laptop without Mojo or AVX-512 vector instructions. About 0.55 GB goes over: stores, text index, two embedding models and four small tools. On 3 of 3 test questions the top 5 matched the desktop, at 2.5 to 2.8 seconds each.
A local chat model on top had mixed results:
The copy is a snapshot, so it must be synced again after notes change. The Wikipedia network and the PDF text store were not copied.
The network is a good memory and an average ranker. Agents use it mostly through the plain gate: recall, read, then walk one step at a time. They get most of their value from being forced to look, and the network adds neighbours the words alone would miss.
If you want to try the idea on your own notes, the cheapest path is the one we took: start with exact search and a recall rule, add vectors, then add links only where you can measure that they help. Our own measurements say that bar is high for a sparse vault, and low for anything dense.
Commenti
Ancora nessun commento.
Accedi per commentare.