A second brain in four steps

An eight-minute introduction to “four steps for giving an AI a memory.” There was almost nothing here to learn, and holding it up against mine produced the sharpest finding in this bucket — my own step four was installed and had never once run.

The four steps

It arranges memory as a four-rung ladder. The point is not the number of rungs but when the human’s hands come off.

Rung	The state	What the human does
1	Explain from scratch every time	Everything
2	Put the material in files and attach them	Attach them
3	The AI finds and reads them itself	Just put them there
4	The AI accumulates and organises on its own	Nothing

Downward, the hands come off. So this is not a ladder of features but a ladder of letting go.

Its closing line is the thesis:

The least replaceable data in the world is data about you.

There is no evidence in it at all

Stated plainly. Zero empirical content.

The only evidence is a comparison demo at the open, and it shows neither the prompt used nor what differed between the two configurations. So it cannot be reproduced.

Memory compounds also appears. That is a metaphor, not a measured result.

→ Fine as a source for understanding the concept, unusable as grounds for a claim about effect. Different jobs.

Held against mine — I was ahead

Rungs two and three were further along on my side, and the rest were things I was already doing.

So it nearly closed as nothing to learn. Then filling the comparison table to the end produced two things.

Finding one — my material was lopsided

The video’s thesis was data about you. I held that ruler against my memory drawers.

All of it was about work. How a tool is used, what accident happened, what rule was set — work, all the way down.

The axis corresponding to “how do I judge things” was entirely empty.

That is a question I would never have asked without the video. Looking only inside my own drawers, they appear well filled. A missing axis is only visible when a ruler comes in from outside.

Finding two — and this one hurt more

I first wrote my step four down as equivalent. It is installed, the structure is right, the code is there.

Then I ran it.

It had never run.

Installed, yes. Code intact. And zero execution history.

So I deleted equivalent and flipped it.

Which is how I learned “I already have that” has three layers

This changed how I make that judgement. One sentence was covering three different states.

Layer	What to ask	How to check
1	Is it there	Does the file exist
2	Does it run	Press it — does output appear
3	Have I ever switched it on	Is there a last-run date

I looked only at layer 1 and said yes. It failed at layer 3.

Three entirely different states, and from the outside they are indistinguishable — a screen showing a file that exists looks the same as a screen showing one that runs.

And it surfaced precisely because there was nothing to learn

One last note.

This was a source with almost nothing to teach. Precisely because there was nothing to take, I filled the comparison table to the end, and filling it exposed the blanks on my side.

With a source full of things to learn, I am busy learning and never look at my own thing.

→ When you use someone else’s thing as a ruler, the ruler does not have to be excellent. It only has to have markings.

The detailed record starts here

The concepts are accurate and the delivery is excellent, and against my setup there is exactly one new adoption. Steps two and three were already ahead; the rest I was already doing. But the biggest return here isn’t the adoption. Holding it up surfaced two things. One is that my material leans entirely toward work, leaving the “data about me” axis empty. The other stings more. My step four was installed and had never once run. I wrote “equivalent” first, then actually ran it and reversed myself.

What it is

A seven-minute-fifty-four introduction. It organises memory into a four-rung ladder with a diagnostic question at each rung. The last line is the theme — “the most irreplaceable data in the world is data about you.”

Three things here are rare for the genre, so I’ll note them first.

• It doesn’t push you to the top rung. Once in the premise and again inside step four, it says “most people probably don’t need this.” It hits the brakes at exactly the step where it introduces its own product.
• Think backwards — decide not what to store but what you’ll ask later. It derives the storage criterion from the output rather than the input, and it is effectively the only thing that keeps a second brain from becoming a junk drawer.
• It grounds the limits of the rules file in the official documentation rather than intuition. It puts up “treated as context, not enforced configuration” and “the more specific and concise, the more consistently followed”, and draws out “it not being obeyed isn’t a bug, it’s a property of this file.” That is a spot where a lot of people misdiagnose the cause, so it earns its keep.

What broke

There is zero demonstration. The only evidence, an A/B in the opening, shows neither the prompt nor the difference in settings, so it can’t be reproduced. Nowhere is there a before/after or a measurement. “Memory compounds” is a metaphor, not a measured result. ⚠ Read it as a conceptual piece and never cite it as evidence for an effect.

Step three breaks its own logic. It says “don’t pour, digest” — and then the largest source of material it proposes is exporting your entire conversation history. That is pouring. The hard part, the digesting procedure, is skipped over with a graph animation. Step three has a proposition and no method.

It never treats the force running the other way. Summarising and compressing are lossy directions. A wrongly compiled memory is harder to fix than the original and poisons every answer after it. ⚠ Compounding applies to losses too, and only the gains are compounded here.

The step-two conclusion is right but not actionable. “Write only pointers” is correct. But the hard parts — what to use as a pointer, when to load it and when to retire it — are all missing. My three-way rule and 45-day retirement exist to fill exactly that gap, and the video has no concept of retirement at all.

And there is zero mention of security. It tells you to gather call recordings, entire note archives and papers into one place while saying nothing about where that folder syncs, what happens if work material gets mixed in, or what it costs in space. In my environment those are real risks.

⚠ Step four is closer to product placement than information. Its only example is his own product, related thumbnails appear immediately after, and the last fifty seconds advertise another product of his. The honest braking is real, but discount step four’s informational value accordingly.

Held against my own setup — and “equivalent” flipped

When I first filled that table in, I was judging by possession. A self-evolution loop exists and the relevant tooling is installed, so I wrote step four down as equivalent. Running it said otherwise. The list, the files and the hooks were all zero.

The lesson is one line — possession is not operation. ⚠ And it doesn’t stop at that one cell. All four “equivalent” verdicts in that table were judged on whether something was installed, not on whether it runs. Until the rest are measured the same way, they are numbers on a ledger.

And the next action isn’t wiring. Apply the video’s own premise — “if nothing hurts right now you don’t need to climb” — and if there was never a moment I missed it, discarding it unwired is also a correct answer. What is needed before deciding is not work but an answer to that question.

Verdict

What	Verdict
Export conversations → fold into the material	adopt — the only new thing. Cumulative usage is 25.8 billion tokens (all machines), and app conversations have never once been folded in
“Will this still be useful in three months”	rejected — duplicate. My three-way rule is already the entry gate. Reopens if the 45-day retirement rate exceeds 30% (a sign the gate is loose)
“Think backwards”	rejected — tautology. Identical to my ② “findable = pointer + reopening condition”. No reopening condition
Rules-file limits as grounds	rejected — kept as a citation. I already have a tool that measures rule compliance, and measurement outranks assertion

The one adoption carries three preconditions. ① separate work material from personal — if they can’t be separated, scrap the item entirely ② originals go under a sync-excluded folder ③ ⚠ never fold the raw export in wholesale — only the digested output. Break that and I am repeating the exact “breaks its own logic” charge I made above.

The baseline being low is also worth writing down

⚠ Being ahead here is not evidence that my structure is good. The counterpart is a seven-minute introduction. It is only evidence that there was nothing to learn from this video.

The real benchmark is another piece in this category. The LLM wiki already covers the concept this video cites at step three, at source level, and the collection automation in the second-brain interview is ahead of my current setup.

The lesson I paid for here is about the comparison table, not the subject — a verdict of “already doing it” is guaranteed wrong if you reach it by checking what’s installed. I pressed one of four cells and it was empty, and I still haven’t pressed the other three.

And ⚠ material with nothing to teach still earns its keep. What this video gave me was not a technique but one unused source of material, and the fact that a cell I believed I owned was a bookkeeping entry.