How to Make AI Remember Your Business Between Sessions

AI tools forget everything between sessions. The fix is a small set of plain text files the AI reads at the start of each one. I keep three: a core file describing the business and how I want the work done, an index listing every other note in one line each, and a running log of what changed. The index is the file that decides whether any of the rest ever gets read. Claude Code loads only the first 200 lines of that index, or the first 25KB, whichever comes first. Everything past that point is dropped, and nothing warns you.

Key takeaways

• Three files cover it: a core file, an index, and a log. No plugin, no database, no subscription.
• The index line is the only text guaranteed to be read. Every other note gets opened only if its line earns it.
• Write each index line to say what the note decides, not what it covers.
• Name files after the rule inside them. A file named after its topic will hide its rule.
• Old notes cause more damage than missing ones, because a wrong fact that looks official gets trusted.

---

Why does AI forget what I told it last week?

Every session starts empty. Anthropic's Claude Code memory documentation states that each session begins with a fresh context window, and that anything meant to survive has to live in files the tool loads at startup.

There are two of those mechanisms. You write one yourself, usually a file called CLAUDE.md that holds your standing instructions. The AI writes the other one, saving notes as it works. Both load at the beginning of every conversation.

So the question is never really whether the AI remembers. It's whether the right file got loaded, and whether the right line inside it got read.

---

Which three files do you actually need?

Three files handle almost everything. I run this setup across 11 client projects and roughly 130 notes, and I've never needed a fourth type.

Keeping them separate matters. The core file answers who and how. The log answers what happened. Mix the two and the core file grows until the useful parts are buried in old status updates.

I've published my full version of this, including the exact structure and the checks I run on it, as a free AI memory setup guide.

---

Why does a correct note still get ignored?

Because saving the note and getting the note read are two different problems, and only one of them is obvious.

Two weeks before I wrote this, I decided to stop putting keyword calls-to-action on my LinkedIn posts. Too salesy, and everyone does it. That decision was saved the same day, with my reasoning, in the right folder.

Then the AI planned my whole content week with that exact call-to-action on five of the seven posts.

The note existed. The note was accurate. Its one-line description said something close to growth strategy notes, follower audit, engagement skill, best-practices pointer. Every word of that was true. It also sounded like background reading, so it never got opened.

The mechanism is documented. Anthropic's memory docs say topic files "are not loaded at startup" and that Claude reads them on demand instead. The index line is the entire basis for that decision. If the line undersells the note, the note may as well not exist.

The fix costs nothing. Rewrite the line so it leads with the decision.

Naming does half the work for you. A file called never-link-personal-profile.md cannot hide its rule. A file called client-social-media.md can, and mine did.

---

How much can the index actually hold?

Less than most people assume. The Claude Code documentation is specific: the first 200 lines of the index file, or the first 25KB, whichever comes first, load at the start of every conversation. Content past that threshold is not loaded at all.

There's research behind that. In Lost in the Middle, Nelson Liu and colleagues at Stanford tested how models use long inputs and found accuracy is highest when the relevant information sits at the very start or the very end. Bury it in the middle of a long document and performance drops sharply. The paper ran in Transactions of the ACL in 2024.

Which is a fairly technical way of saying what every writer already knows. Put the important thing first, and keep the document short enough that there isn't much of a middle.

---

Why are old notes more dangerous than missing ones?

A missing note produces a question. A stale note produces a confident wrong answer.

My own core file carried a revenue figure that was off by more than double. Nobody planted it there. I'd said the number out loud months earlier as a rough guess in conversation, it got written down, and it sat in the file looking exactly like a verified fact ever since.

Contradictions are the other half of this. The documentation is blunt about what happens when two of your rules disagree: Claude "may pick one arbitrarily." You won't get an error. You'll get whichever answer it landed on, delivered with the same confidence as the right one.

So retire notes rather than deleting them. Mark the old one superseded, add the date, and point it at whatever replaced it. Deleting destroys the reasoning, and six months later you re-litigate a decision you already made.

Four checks, once a month, catch nearly all of it:

• Notes older than six months that nothing has updated
• Notes that contradict each other or contradict the core file
• Numbers you originally guessed rather than looked up
• Notes nothing has referenced since the day you wrote them

---

What will this not fix?

Written rules are guidance. They aren't a switch.

Anthropic says it directly in its own documentation: Claude treats these files as "context, not enforced configuration," and there is no guarantee of strict compliance, especially where instructions are vague or conflict with each other.

The bug reports say the same thing louder. In January 2026 a user opened issue 17530 on Anthropic's public tracker, reporting that explicit instructions were ignored in roughly half of all sessions. He had written the rules in capitals, labelled one a hard rule, and corrected the AI six separate times in the same conversation. It cost him 55 minutes on a five-minute fix. Anthropic closed the issue as not planned.

Worth being precise about what that is. One user's report, not a study and not Anthropic's own measurement. But it matches the documentation and it matches what I see.

Good notes raise the odds that the right thing happens. They don't make it certain. That's an honest limit, and I'd rather say it than sell a system that quietly fails on you.

Everything else on this page is the cheap 90 percent. Thirty minutes of setup, three files, and a habit of writing the index line properly.

---

Frequently asked questions

Do I need a plugin or a paid memory tool?

No. This runs on plain markdown files in a folder. I looked at the popular memory plugins before deciding, and several have logged bugs where a routine update wiped saved history or corrupted the search index. Text files have no database to corrupt, and you can read every one of them yourself.

How long does the initial setup take?

About thirty minutes for the first project. Write the core file describing your business and how you want work done, create an empty index, and start the log with today. The notes accumulate on their own after that, because you add one each time you correct the AI on something.

Does this work with ChatGPT or only Claude Code?

The structure works with any AI tool. The 200-line loading limit is specific to Claude Code auto memory. With ChatGPT projects or a custom GPT, you paste or attach the same three files instead. The index line rule matters even more there, because you are choosing by hand what to include.

How often should I clean up the notes?

Once a month is enough for most people. Check four things: notes older than six months, notes that contradict each other, numbers you originally guessed, and notes nothing has referenced since you wrote them. Retire what is out of date by marking it superseded with the date rather than deleting it.

Where should I start if I only do one thing?

Open your index and read it line by line. For each one, ask whether it tells you what the note decides or only what it covers. Every line that names a topic is hiding whatever sits inside it. That pass took me an afternoon and turned up 19 rules my AI had never once read. If you want a head start on the core file, the context file template covers what belongs in it.