FIELD NOTES · MEMORY SYSTEMS · N = 1

I studied the first brain.
Then I built a second one.

Before cybersecurity I was pre-med at Loyola. I got the neuroscience degree, took the MCAT, and my scores were not great. Ten years later I am building a second brain out of markdown files and AI agents, and the two fields keep explaining each other. These are the notes, with figures you can poke at.

1Capture works because encoding is cheap

The hippocampus does not file memories neatly. It grabs a fast, sloppy trace and moves on, the sorting happens later, mostly while you sleep. My second brain's inbox obeys the same rule: one line, timestamped, exact wording, no organizing allowed at capture time. The biology figured this out first. If saving a memory required filing it correctly, you would save almost nothing.

Fig. 1 A memory is a pattern, not a file. Move your cursor over the plate to excite a neuron and watch the trace spread. Cells that fire together wire together (Hebb, 1949).

2Forgetting is the default, review is the override

In 1885 Hermann Ebbinghaus memorized thousands of nonsense syllables and tested himself for years. The result is the forgetting curve: retention collapses within days unless you review. Each review resets the curve to full strength and makes the next decay slower, which is the spacing effect, and it is the entire scientific case for flashcards.

It is also why my weekly review exists. Skimming last week's notes is not discipline theater, it is the override switch on a curve that is otherwise headed to zero.

Fig. 2 The Ebbinghaus forgetting curve, R = e−t/S. Click anywhere on the plot to schedule a review on that day; each one resets retention and raises stability ×1.9. Try earning a flat curve with the fewest reviews.

3Your context window is a working memory

George Miller measured it in 1956: humans hold about seven chunks in mind, plus or minus two. The trick that beats the limit is chunking. You do not hold more items, you hold bigger ones, "FBI CIA IRS" is nine letters or three chunks depending on how you encode it.

Claude has the same constraint, truly the same shape. A context window is a working memory, a bloated one gets dumber, and the fix is chunking there too. A skill file or a CLAUDE.md is one handle that expands to a whole procedure, three letters that stand for nine.

press start
Fig. 3 Digit span, the classic working memory test. Digits flash one at a time, type them back. The sequence grows until you miss twice. Most people land between 5 and 9 (Miller, 1956).

4The homology table

Same problems, different substrate. The mapping is closer than it has any right to be!

First brainSecond brainShared problem
Hippocampal encodingan inbox capture linesave first, organize later
Sleep consolidationthe weekly reviewkeep what matters, drop the rest
Ebbinghaus decaynotes you never reopenunused memories fade
Long-term potentiationresurfacing notes on a cadenceuse strengthens the trace
Working memory, 7 ± 2 chunksthe context windowlimited slots
Chunkingskills and CLAUDE.md filesone handle, a whole procedure
An engram, the physical tracea markdown filewhere the memory actually lives

The degree I thought I wasted turned out to be the design spec. The second brain itself is nothing exotic, plain files and one loop: capture, sort, decide. The neuroscience just tells you why the loop has to look like that.