The AI Memory Setup That Survives

Every AI conversation starts from nothing. You explain your business, your clients, your preferences, and the moment you close the window it is all gone. Next session you start again.

Most people fix this by writing notes for the AI to read. That works, right up until it quietly stops working, and the way it fails is not obvious.

---

Start with three files, not one

The instinct is to keep one big file with everything in it. That file goes stale within a few weeks and then you stop trusting any of it.

Split it into three instead, based on how fast the information changes.

Keeping them apart is the whole trick. When current state and permanent rules live in the same file, updating one means touching the other, and eventually nobody trusts either.

---

The contents page is the part that actually matters

If you have more than a handful of notes you need a contents page listing them, one line each. Your AI reads that page at the start of every session. It opens an individual note only if that one line makes the note sound worth opening.

This is where my system failed, and it took two weeks to notice.

I had decided to stop putting comment prompts on my LinkedIn posts. That decision was saved correctly, the same day, with my reasoning. Two weeks later my AI planned a full week of posts with those prompts on five of seven.

Write each line for retrieval, not description

Four things make a line work:

• Lead with the rule, not the topic. A topic gets skipped. A decision gets read.
• Name what would go wrong. A consequence is memorable. A subject is not.
• Say when it matters. Add a trigger like "read before writing any client email." Your AI needs to know when a note becomes relevant, not just what it covers.
• Use the words you will actually type. This is the one people miss. Your note might say "growth strategy" while you type "plan my posts." Those never connect.

Name files after the rule, not the subject

Here is a shortcut that costs nothing. When I audited my projects, every note I had named after its rule was fine. Every note named after a topic was hiding something.

A file called never-link-personal-profile cannot hide its rule. A file called client-social-media can, and mine did.

---

There is a size limit nobody mentions

Only the first 200 lines, or the first 25KB, of your contents page is read at the start of a session. Whichever comes first.

---

Old notes are more dangerous than missing ones

This surprised me. When I researched what goes wrong with these systems, the most common failure was not information being absent. It was old information being confidently used.

A note that was true in June and quietly wrong in August is worse than no note, because it gets trusted.

Four things to check every month or so:

• Anything untouched for three months that describes something changeable. A status, a price, a queue, someone's role. Facts age at different speeds. "Our brand voice is direct" lasts years. "The site is mid-rebuild" lasts weeks.
• Two notes that disagree. Flag both. Never pick one quietly, because that is how a wrong fact becomes the permanent record.
• A note that contradicts your main instructions file. That is a live bug, not a wording problem. Mine told my AI to publish through a service I do not even have an account with.
• Notes nothing points to. If no line in your contents page meaningfully mentions it, it does not exist.

---

Retire notes, do not delete them

When something is genuinely dead, resist deleting it. A deleted rule can resurface later from an old file or an old conversation and get applied again without whatever replaced it.

---

Why plain text beats the proper tools

I assumed I was behind. There are real tools built for this by people who have thought about it much harder than I have, and I am running text files in folders.

So I ran two separate deep research passes to find out what I was missing. Both came back with the same answer: keep the text files.

The reason is not that the tools are badly made. It is what happens when they break.

• One popular tool permanently deleted 312 messages of someone's work during a routine automatic update. The log said "recovered," so he did not notice for days.
• Another corrupted its own search index, and every lookup failed permanently afterward.
• A third silently broke all its searches while continuing to run and charge for it.
• Installing one version quietly switched off the built-in memory on a machine holding 75 existing notes, which stayed on disk but became invisible.

Text files in a folder cannot do any of that. There is no database to corrupt and nothing that updates itself behind your back. If something goes wrong you open a file and read it.

Add version history through something like Git and you can undo any change, see exactly what changed and when, and recover anything.

---

The thing this will not fix

Worth being straight about. A written rule is guidance, not a control.

Anthropic's own documentation says these files are treated as context rather than enforced configuration. Their public issue tracker backs that up: issue 17530, opened in January 2026, reports instructions being ignored in roughly half of all sessions, with the rules written in capitals, labelled a hard rule, and corrected six separate times in the same conversation. It is one user's report rather than a study, and Anthropic closed it as not planned.

---

Getting started this week

You do not need the whole thing on day one.

1. Make one file for rules and one for current state.
2. Write down the five things you are tired of re-explaining.
3. Name each note after its rule, not its subject.
4. When you have more than about eight notes, add the contents page and write every line so it leads with the rule.
5. Set a reminder for a month out to read back through and ask one question about each note: is this still true?