Claude Code Has Been Remembering Your Preferences Since February, but It Only Loads the First 200 Lines

In any Claude Code project, type /memory, and you'll see an auto memory folder that has existed since February. It's enabled by default in local sessions. Many people have never opened it.

Claude Code is Anthropic's coding agent, and it can read project directories. Readers who don't write code can think of it as an assistant that rummages through folders for information and also takes notes. Being on by default is pretty neat. You forget; it doesn't.

First, see what it remembers

The steps aren't complicated:

1. Open any project and type /memory
2. Check the auto memory switch. It's on by default for local sessions
3. Choose to open the auto memory folder
4. Read MEMORY.md first. It's an index, one memory per line
5. Then look at the files. Each memory is a file, with types user, feedback, project, and reference
6. Tell Claude remember: always use pnpm, not npm, and it will save a new memory
7. You can edit or delete incorrect memories directly. It's pure Markdown

The original tweet includes a demo video:

The key limitation is loading: each session only reads the first 200 lines or 25KB of MEMORY.md. Claude opens topic files as needed. Memories stay until a person or Claude edits or deletes them. Original tweet link: plusxbt.

The folder isn't complicated; maintenance is

This approach isn't only for auto memory. @Bober_smart wrote up a Claude Second Brain & Folder Workflow. The method is a flat Markdown folder plus a CLAUDE.md index. When using Claude Code, CLAUDE.md automatically goes into the system prompt at startup. It's like putting the map on the table first.

Keep the directory simple: notes, people, projects, with MEMORY.md, LEARNINGS.md, and decisions.md at the root. The rules are simple too: one file per topic, and filenames use lowercase hyphens. Use @ references for files that need to stay in context, but don't overuse them. @Bober_smart's judgment is that proprietary formats from Obsidian and Notion add an intermediate layer for agents, while plain text has fewer walls. That judgment also holds for ordinary users. Files are on the hard drive, readable and editable, and you don't need to learn a database first.

The problem is memory quality. In a reply, John Rood offered a sharper judgment: this folder stores conclusions. The evidence isn't in it. A line like "use X" has no date and no reason, so it won't expire on its own. He built memoryrouter for this, so decisions, reasons, and dates travel together. This criticism is worth remembering. Default memory saves you from repeating yourself, but it also turns context into invisible debt.

Someone else reminded people about the 200-line limit. Once MEMORY.md goes over that, later lines stop loading, and Claude quietly forgets them. Clean it up every few weeks. Another person asked how many people actually read what Claude has stored before sharing a project folder. The question isn't fancy, but it's practical. Memories are pure Markdown and travel with the folder. Opening them for a look before sharing is easier than guessing afterward why it did something.

There's no install script or complex configuration. It's a set of Markdown files, an index, and a discipline. The real cost isn't in setting it up; it's whether you're willing to open the folder and take a look each time the memory gets dirty. Start by typing /memory.