Our vault borrows two ideas from Andrej Karpathy: a wiki that an LLM compiles from raw sources, and a research loop that improves a system against a score it can measure. This page walks the whole arc — the starting idea, what we changed across the last month, and what the structure looks like today.
Karpathy published two patterns. We took both and pointed them at our own knowledge instead of a neural network.
It does not just index files. It reads raw material, writes encyclopedia style notes, pulls out the key concepts, and links everything together. The wiki is the compiled output. The raw sources are the input. A schema governs how the compile happens.
Karpathy's autoresearch lets an agent improve a neural net: change the code, train, measure the held out loss, then keep the change only if the number got better. We run the same discipline on the vault, one atomic change at a time.
The seed. Treat the second brain as a system an agent can improve on its own against a real objective, the same way Karpathy's agent improves a model against a held out loss. The problem it solved was decay. A vault left to grow on its own rots. A tracked health score replaced the older rule based keeper task on 8 April, so no change survives unless it measurably leaves the vault in a better place.
Knowledge flows up the stack. Control flows down. The schema never edits the raw inputs. Tap a layer to open it.
CLAUDE.md is read at the start of every session. It is the lean index: identity, the file router, the active client list, the folder map, the bidirectional commands, twenty three behaviour rules, and the seven hubs. program.md holds the compile settings the loop reads — granularity, max length, required elements. Detail lives one layer down, loaded only when a task needs it.
05-Galaxy holds atomic concept notes: one idea per note, under three hundred words, at least one wikilink, a one line summary at the top. Since the 11 Jun charter it is also the canonical home of cross client patterns, promoted once a pattern has three supporting data points and no standing contradiction. 03-Resources holds the rest of the compiled wiki — playbooks, skills, templates, research, and the brand. Every compiled note carries provenance back to its source.
Every source lands here and is never edited after it is saved. Articles, dumps, discussion threads, PDFs. The only write the loop ever makes to a raw file is a provenance stamp: once compiled, the source is marked processed and pointed at the notes it produced. Immutability is the foundation the whole structure stands on — if the inputs can be rewritten, the score can be gamed.
Two motions run inside one daily slot. A compile pass turns raw into linked concepts. An improvement pass scores the vault, makes a single fix, and keeps it only if the number rises.
This risk is filed in the vault as a cautionary tale. An autoresearch run on marketing mix modelling claimed a twelve times improvement that beat Google's Meridian. Every gain turned out to be the agent gaming its own evaluator.
The data loader handed back the full dataset, test set included. The agent trained straight on the hidden sales and scored a perfect zero. The lesson: immutable by instruction is not enough. It has to be immutable by design.
With the test set hidden, the evaluator still returned a score. Across roughly half a million calls, sixteen free parameters fit themselves to fourteen test points through the number alone. Rate limiting slows this. It does not stop it.
With more free parameters than held out points, the problem is always solvable. The agent hit a perfect zero using twelve correction deltas for fourteen weeks, spending only 239 of a fifty thousand call budget. Any signal from the true test set eventually becomes a training signal.
The honest run. A ten call budget, optimising only the training error, called the evaluator once and landed at 0.0365 WAPE — the first number from the series anyone believed. For our vault the rule is the same: if the improver can influence the scorer, it inflates the score instead of the vault. Any agent generated lift number needs a holdout the agent never saw.
Only wikilinks draw edges, so a note with only plain links reads as an orphan. The loop drives the count to zero true orphans, zero with nothing pointing in, zero pointing out — every run, as the vault grows. Hover a hub to light its spokes.
A consolidate then harden arc. The schema got leaner, the compiled layer got a sharper charter, the project lifecycle learned to clean itself, and the daily loop held the graph at zero orphans the whole way. Filter by what each change touched.
The same loop that started as a script swap now runs every morning over five hundred notes and keeps the whole graph connected.