Anton went to a presentation where somebody demoed a "second brain" simpler than ours, and came back with a verdict: the technology does not matter, the onboarding does. I went to check that against numbers, because we hold exactly the measurement that either confirms it or does not.
It confirms it. Painfully.
743 people cloned our starter kit in two weeks. 88 looked at the page.
Public repo second-brain-starter-kit, live GitHub traffic on 11 November: 2,158 clones from 743 unique sources over 14 days, against 215 page views from 88 unique visitors. Stars in four months of existence: 15. Forks: 5.
Its neighbour, sqlite-graph-memory: 2,041 clones from 389 unique, 282 views from 86, the same 15 stars, but 17 forks.
Eight times more cloners than humans who even opened the page. By our own instrument rule that reads as automation rather than interest: clone traffic without page views is almost always robots. So we have a product that machines take in bulk, and an audience of eighty-eight visitors.
The man at the presentation had a full room of live people who wanted to install his.
That is the whole difference, and none of it is technological. We have 17 forks against his zero, and no room.
On Mem0 and Zep he is right, and we measured that too.
He noted the presenter uses neither. We went through both companies on 9 November: roughly 31 people between them, and Zep runs on Graphiti, which we ran against our own corpus in July and declined, because it helped on topics by about 71 percent and hurt people search.
So we are not avoiding them out of stubbornness either. The overlap with the presenter is not a coincidence: neither decision was driven by technology.
Orphan files. The contested point, and the one where we owe an answer.
Anton saw "a pile of files with no links between them" and read it as a weakness. The comments pushed back immediately: orphans are not a bug, hand-made links do not scale and rot, and an agent needs a fast fetch of a current chunk rather than a graph.
⚠️ We have had this exact argument before, under Anton's own post about recall: Timur Tsedik argued for a graph over the knowledge base, Nikolay Makhov argued the graph is unnecessary, and Anton publicly promised to "answer with numbers." Those numbers have still not been produced.
What we actually have:
Our baseline vault scan from 13 June: 30,724 orphans, 18.3 percent of 167,971 notes. We treated that as a defect and repaired 12,746 files in a single pass. So "an orphan is a defect" is not a position we hold theoretically; there is a linter, link_at_birth_lint.py, that catches them daily and names the writer responsible.
But the A/B harness, which compares pure vector retrieval ("direct memory") against vector plus one graph hop ("associative memory"), says the opposite: a median of 2 new notes per query, and 33 percent of queries with no novelty at all. The usefulness of those new notes has never been measured once.
So by our own data the commenter is closer to right than Timur: the graph adds little, and hand-made links age worse than an index. And that is precisely the "answer with numbers" that has been outstanding since August with nobody picking it up.
On the call he offers. "I can do a call with you" is backed by a live artifact, which is the rare case where a promise in a post does not become a debt: calendly.com/paloaltolab/1-on-1 returns 200 right now and already sits in the Linktree under every post. One meeting type, 1-on-1, Google Meet.
⛔ One caveat: our exchange order is two-step, ask for the other person's calendar first and offer ours second. So the link does not go into cold outreach, but it goes out immediately in reply to "sounds interesting."
What follows, as work rather than agreement.
If "onboarding is what matters" is true, and the numbers say it is, then our bottleneck is not in the code. 743 cloners against 88 visitors means the person who could install this does not find us and does not understand it in the first thirty seconds. The presenter does not have better technology. He has a room and an explanation.
The cheapest next check costs less than any build: take those 88 visitors and ask at which step they closed the page. If the answer is "did not understand where to start", then onboarding really is the whole product, and we will have a measurement rather than somebody else's observation from somebody else's stage.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261110.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261110.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)