A new ICM layer?? (Gifts for everyone at the end)

I think I cracked the code on why we are all having a hard time with this second brain stuff.

The real problem is that we are saving information that we think will be relevant to future endeavors, not just the stuff that is presently needed. Thus, we need rules that work in many (read "all") workspaces to make the information usable/searchable etc. These hypothetical rules would likely be more constraining than the rules that an actual use case would require.

The second issue is that ICM is (correctly) very broad. Its principles map extremely well to the constraints of a task at hand, but poorly to undefined future tasks (like the goofballs making ICMs for imaginary problems)

Possible solution: We stop saving unnecessary info and let our desired function determine the format (within ICM L3 Principles).

That solution sucks, I still want to save info I don't need today!!

I found a better solution...

(If you just want the solution skip to the last paragraph. If you want some cool info on why your second brain doesn’t work, read on.)

In practice, our L3 files are almost all SOPs. But the way an SOP is structured doesn't work for the way externally-authored source material should be kept for citation or lookup (which btw = split into its natural segments and indexed).

This is why Tiago Forte’s PARA system doesn’t work, it’s built for helping a human decide what to do next, not have access to the relevant info.

Google’s OKF fails because it's a format for shipping knowledge between tools/organizations, not the method for finding that info.

Karpathy’s wiki fails because it defines only the loop (ingest / query / lint), so retrieval quality is only as good as the agents improvisation, its unreliable.

RAG and vector databases are for bulk data, not a second brain. For instance, a million support tickets is a pile to query, not a knowledge base.

Like our boy Jake likes to say (shout out Jake!), the principles for darn near everything have been around for decades. We're just applying timeless principles to an unrefined tool. The principle needed here is called faceted classification, developed by a librarian named S.R. Ranganathan in 1933, decades before we had a database to apply it to.

Alright, sorry for the clickbait title... the fundamental difference between an SOP and a 2nd brain isn't the layer.

The difference is that a 2nd brain is a different type of CONTENT STRUCTURE than what we ICMers may be accustomed to. It's also different than real life (!!): physical data (books etc.) can only exist in one place at a time, so systems like the Dewey Decimal system are as good as it gets. The trade off is that a book might cover two vastly different topics at the same time, but it can only go on one shelf. The benefit of digital data is that many facets are simultaneously searchable. For instance on Amazon each product is tagged: size, color, brand, and price etc. are independent lists you combine at search time and it filters out anything that doesnt match what you need.

Ranganathan and the standards bodies after him (ANSI/NISO Z39.19, ISO 25964) needed a second half to complete the system: a list of aliases, formally called entry vocabulary, to find the correct files. For instance, if you searched "dark red" but the content was stored under a different facet value like "maroon," you got nothing, unless a human had sat down and entered "dark red, see maroon." By hand... in advance... for EVERY single term anyone might use.

Now, in the year of our lord, 2026, the matcher is an LLM. It gets from dark red to maroon no problem, because it "understands" what the words mean instead of which characters they contain. So the half of the system that was pure upkeep is now irrelevant (thank God). What's left is a controlled vocab list the AI can map any relevant question onto.

Ranganathan’s faceted classification on top of ICM principles (SSOT, Filesystem is the state machine, load only what the step needs, etc.) with some inspiration from OKF and Karpathy led me to this set of principles that I think achieves the “generalizable rules” goal a 2nd brain requires:

1. Controlled vocabulary + entry vocabulary (now the LLMs job); ANSI/NISO Z39.19, ISO 25964

2. Atomic (small) notes; Luhmann's Zettelkasten, this is also a strong ICM principle.

3. Postcoordinate indexing; Mortimer Taube, Uniterm, ~1951

4. Literary warrant (create a category only when note quantity justifies it); E. Wyndham Hulme, ~1911

5. Term specificity / IDF (a term applied to everything specifies nothing); Karen Spärck Jones, 1972

6. Precision vs recall (known-item vs exploratory ie. ”find this fact” vs “find info on this topic”); Cleverdon, Cranfield experiments, 1960s

7. Associative trails over hierarchy; Vannevar Bush, "As We May Think," 1945

8. Google’s OKF (the lifecycle and trust fields I hadn't thought to build)

9. Karpathy (agent-maintained markdown wiki with ingest / query / lint.)

And the only principle I’m actually contributing here:

10. Save the non obvious, not every fact (AI doesn’t need the reminder “use proper punctuation”)

Please note that even Google, Karpathy and Forte are just syntheses of what came before. We stand on the shoulders of giants.

GIFT HERE:

This stuff is kind of a lot, so I have a gift for you. I created a system to audit my workspace and make sure it's running the way ICM says it should, AND to handle the 3 content structures I've identified. You get 4 L3 docs (plus everything else):

a general L3 audit that also determines the file type below:
one for SOP type L3 files
one for source material (the stuff you cite but never rewrite)
one for knowledge bases (the actual 2nd brain)

It's zipped and attached below. Drop it into your workspace, point your agent at audit-root.md and tell it to audit or create your 2nd brain. Its awesome.

P.S. audit L0 second ;)

20 comments

A new ICM layer?? (Gifts for everyone at the end)