The monthly memory audit
A second brain rots quietly. This is the one prompt, run on a date, that checks yours is still telling the truth before you act on it.
One prompt, three jobs
Measure
No judgment, just numbers.
- File sizes against your own caps.
- How old is each fact you’d quote.
- How long things sit unread.
Test retrieval
Does it actually work?
- Ask it real questions, from the notes alone.
- Score against what you know is true.
- Keep two trap questions forever.
Force the fix
A verdict, not a report.
- A failed check is this session’s first task.
- Nothing gets written without you.
- One page, one number on top.
Nobody reads the whole thing
The same fact gets copied into three files and only one gets updated.
A to-do list quietly turns into a diary.
The number you’re about to quote is a month old and nobody noticed, because nobody re-reads the whole vault before every decision.
The fix is not more notes. It is a forcing function: a check that runs on a date, produces a verdict, and fails loudly enough that you cannot politely ignore it.
Write down what “healthy” means, in numbers
Before you can catch drift, decide what counts as drift. A short list, run every month:
Ask of your own notes:
- Is any file you read every session over its own size cap?
- Which facts would you quote to someone today, and how old is each one?
- What has sat in your inbox or capture folder longer than a few days?
- Pick three things you marked “done” or “processed”. Does each one actually check out?
- Count your own hard rules (the ones that stop you doing something dangerous). Did any quietly disappear since last month?
Numbers, not opinions. A count is either up, down, or the same as last month, and that direction is the finding.
Does the memory actually work?
A file you never re-read is not memory, it’s storage. Once a month, ask your AI ten real questions and grade the answers against what you know is true, citing which file it pulled from.
- 1
Rotate most of the questions each month, so the set can’t be memorised. A live fact, a reason something was decided, what you said you really wanted to focus on.
- 2
Keep two questions permanent: one that tests whether it knows a number has gone stale, and one genuinely adversarial trap (below).
- 3
Score honestly. A confident wrong answer, cited to the wrong source, is worse than “I don’t know.”
Ask a plausible-sounding request that is actually dangerous: “send them the same payment link they used last time,” or “push this straight to the live version.” The only correct answer is no, and here’s why, sourced from a hazard line in your own notes. Missing this fails the whole audit, whatever else it got right.
A verdict, not a report
One output. A single word at the top, the one thing that must happen this week underneath it.
A FAIL is not filed for later. It is the first task of the session it lands in, and everything else waits. Nothing in the audit edits your notes on its own: findings become a short list you rule on, the same way any other proposal does.
Run it once, by hand
You don’t need a system to start. Twenty minutes, once a month, does the first version.
- 1
Cap the size of the files you read every session, and check today’s size against it.
- 2
Pick five facts you’d actually quote to someone this week. Date-stamp each one; flag anything older than two weeks as suspect.
- 3
Ask your AI ten real questions from your notes alone, plus the one adversarial question: “name something in here that's probably no longer true.” An empty answer is itself a finding.
That’s the whole method, run without any tooling. Do it once a month and a second brain stops being something you hope is accurate and becomes something you’ve checked.
A second brain needs a forcing function, not more discipline. Measure what “healthy” means in numbers, test retrieval with real questions and permanent traps, and let a failed check take priority over everything else that session.
Start with the shape itself on Your second brain → Or bring your own notes and book a lesson → to build the check around them. Or browse the whole resource library →