Claude Code forgets everything each session. How do you make it remember?

Last Tuesday you explained the supplier

Last Tuesday you spent ten minutes on it. You'd switched suppliers. The old one raised prices. The new one delivers on Thursdays. Claude got it. It even rewrote your ordering checklist.

Today you ask for this week's order list. It builds the whole thing around the old supplier.

You did have that talk. So is it broken? Or am I using it wrong?

Neither. That talk happened in a session that ended.

Why talking to it louder won't work

Each new session starts with a clean slate. The memory page in the official docs says so plainly. Most people find out the hard way, by repeating themselves.

So they repeat themselves harder. Long rules at the top of every chat. Capital letters. It helps for a day, and then a long session gets summed up and the rules fall out of the summary.

The fix isn't more words in the chat. It is words in a file that loads every time.

A notebook by the bed

Picture a helper who wakes up every morning remembering nothing. But next to the bed is a notebook. The first page is a table of contents, one line per topic. "Supplier changed in March, see page 12."

Each morning the helper reads that first page. It only turns to page 12 when the day's work is about the supplier. Each evening, if it learned something worth keeping, it writes a new page and adds one line to the front.

That is how auto memory works. The front page is a file called MEMORY.md. The pages are small note files, one topic each. Claude writes them when you say "remember that...", and it skips things it could work out from your files.

Three limits matter. Only the top 200 lines of the index get read, so keep it short. The notebook stays in one house, because memory lives on one computer. And nothing in it updates itself. A note that was true in March is still saying the same thing in October.

The book's note format has four parts: what, why, how to apply and the date it was checked. The "why" is the one people skip. Without it, a later session can't tell if the rule still makes sense.

Just want yesterday's conversation back?

Sometimes you do not need memory. You need the old chat. Claude Code saves every session as a file, so you can reopen one. The sessions page lists the ways:

Two things to know. Start Claude Code in the same folder you worked in, because sessions are stored per project folder. And transcripts are cleaned up after 30 days by default.

A resumed session is still not memory. If the old chat was summed up to make room, the summary is what comes back, not every word. For a fact you will need next month, a note beats a transcript.

For long jobs, keep a progress file

Memory notes are for facts. For a big job that runs over several sessions, the book uses a third tool: a plain file called PROGRESS.md in the project folder.

The long-job prompt in the book tells Claude to work in batches. After every batch, it updates PROGRESS.md with the files finished, the files still to do, and anything it was unsure about. If the session stops for any reason, the next one starts by reading that file and picks up from the first file still to do. It never redoes a file marked as finished.

My own note from the day I learned this: assume a long job will stop halfway. Usage limits, a full context window or a crash will end it for reasons that have nothing to do with the work.

Three times my assistant got it wrong

By the fall of 2026, my assistant's memory held about 200 notes behind one index. The rules I've stated firmly carry a stop sign there. Before the index, I gave the same corrections again and again. After it, they stopped repeating.

Then it taught me that a stale note is not my plan. In July 2026 it wrote a profile of how I make decisions and said I stall before publishing. It got that from old notes listing my books as unfinished work. All five were already published. I added a rule: confirm what is true now before drawing a conclusion from old notes.

On September 8, 2026, it acted strangely on my work computer. Its memory folder was on the wrong drive. Only one skill was installed instead of the full set. None of the four scheduled jobs existed. The setup script had failed without a word.

The same day I found out why my quote skill had sat unused for weeks. It lived in a folder another AI tool uses, a folder Claude Code doesn't load. No error. Claude simply didn't know it existed. If a skill never fires, check where it lives before you rewrite it.

The chapter that gives it a memory

This is the problem Claude Code for Non-Programmers was written for. The book is independent and is not affiliated with Anthropic.

Memory That Survives gives you the note template, the index format, five memory rules for your CLAUDE.md, a five-minute weekly review and a new-machine checklist.

CLAUDE.md: The Instruction Sheet holds the handbook half, the things that rarely change.

Skills: Teaching It a Job Once shows where skills must live to load. The official skills page lists the folders.