CLAUDE.md and memory: how the AI remembers us

This is a follow-up to the previous post — Building a Personal Knowledge Base on Obsidian + AI: My Approach. If we don’t have our own vault with PARA and a basic CLAUDE.md yet — start there, otherwise this will be like reading a recipe with no pot.

Why the AI “forgets us” without CLAUDE.md

When we open regular ChatGPT in the browser — there’s a history of past chats. That’s convenient, but it’s not memory about us. It’s just an archive of conversations. Every new chat starts from scratch: the AI doesn’t know our name, our projects, or what language we want answers in.

Claude Code in VS Code works differently. It reads files. And among all the files in our folder there’s one special one — CLAUDE.md in the root of our vault. The AI reads it every single time we start a new session. This is not “conversation history.” This is an instruction on how to work with us.

In plain words: CLAUDE.md is the passport of our system. Without it, the AI behaves like a new intern we have to brief from scratch every single time.

7 categories — what to write in CLAUDE.md

We can write whatever we want — but there’s a battle-tested 7-block scaffold. That covers 90% of cases.

Category	What goes here
Who I am	Name or handle, role, city, language. One or two lines.
Response style	English, no fluff, no intros like “Of course!”, markdown, short.
Vault structure	List of folders and what lives in each. The AI uses this to know where to put new files.
Working rules	Don’t touch .obsidian/. No deletions without confirmation. Changes in 2+ files → show the plan first.
Security and passwords	Passwords, access keys, secrets — never store them in the vault and never paste them into AI chat. Keep them in a separate password manager (e.g. KeePass). State explicitly in CLAUDE.md: “don’t ask for passwords, don’t store them in files, don’t show them in responses.” This is the most critical rule.
My projects	Active projects with a short context. 2-3 lines each.
My preferences	How we like our responses formatted, which tools we use, what to avoid.
Forbidden topics	If there’s data the AI should never work with at all — name the folders and files.

Each category — 3-10 lines. The total size of CLAUDE.md should stay within 100-200 lines. If the file grows beyond that — pull parts of it into separate files and just link to them.

Auto-memory — what it is and how it works

Besides CLAUDE.md, Claude Code has a second system called memory (sometimes called “auto-memory” or “persistent memory”). This is the place where the AI itself writes down important things worth remembering between sessions.

Example: we once told it — “for meeting notes I use a template with Date / Attendees / Decisions / Actions.” The AI can record this in memory as a fact. Next time we ask it to create a meeting note, it’ll grab our template immediately, no clarifications needed.

Technically it’s a separate file (usually MEMORY.md in a special system folder outside the vault). We can open it, read it, edit it by hand. But mostly — the AI manages it on its own.

The difference is simple:

• CLAUDE.md — we write by hand. These are our rules.
• memory — the AI writes itself. These are its notes about us.

Together they create the effect of “an AI that knows us.” Not magic — just two files that get read on startup.

How to make memory actually work

By default the AI isn’t eager to write into memory. We need to tell it: “this is worth remembering.” A few tricks:

• Explicit command. “Remember: my Weekly Review is every Friday at 17:00.” The AI will confirm and add it.
• Feedback. If the AI did something wrong — say so. “Don’t use underscores for italics, my messenger breaks them.” It’ll remember.
• Weekly review. Once a week we open the memory file and throw out what’s outdated. If a project wrapped up — remove it from memory. Otherwise the AI keeps thinking it’s still relevant.

3 tests — is memory actually working

How to tell whether the system is alive, not just “files sitting there”:

Test 1. We open a new chat in Claude Code (not a continuation, a fresh one) and ask: “What do you know about me?” The AI should respond concretely — name, language, key projects. If the answer is vague (“I can help with various things”) — CLAUDE.md isn’t being read. We check that the file sits in the root of the vault and that the vault folder is open in VS Code.

Test 2. We give the AI a task that breaks one of our rules. For example, we ask it to make a change in 3 files. If CLAUDE.md says “before changes in 2+ files — show the plan” — the AI should stop and show the plan, not just start working. If it just starts — the rule is being ignored or it’s worded unclearly.

Test 3. We tell the AI to remember something specific. “Remember: my Weekly Review is every Friday at 17:00.” Close VS Code, open it again, in a new chat ask: “When is my Weekly Review?” It should say “Friday, 17:00.” If it doesn’t remember — memory isn’t being written, check the settings.

Common mistakes

1. A CLAUDE.md 500 lines long. If the file is that big — the AI reads it slowly and can ignore parts. We split it by category, pull details into separate files under 99_Meta/, leave only references in CLAUDE.md.

2. Conflict between CLAUDE.md and memory. It happens: CLAUDE.md says “be brief,” memory has “give detailed explanations” written into it. The CLAUDE.md rule wins — it has priority. But it’s better to periodically re-read memory and toss out conflicting entries.

3. Not updating for months. Projects change. If CLAUDE.md mentions a project we closed half a year ago — the AI will keep referring to it. Once a month — a quick review and update.

4. Recovery after a crash. If something breaks and the AI behaves weirdly — don’t panic. Just open CLAUDE.md by hand, make sure the file is intact, readable, with no stray markdown errors. In 9 cases out of 10 the problem is right there.

What’s next

Now we have a vault, the AI knows who we are, and it remembers our rules. The next logical step is to make sure this system doesn’t disappear if our laptop falls into a lake. Backup to the internet, privately, no terror.

If we don’t want to miss it — follow me on LinkedIn, I announce each next piece there.