How to Give Claude Perfect Memory: Steal My 3 Layer System

Clean up your memory

Go to Settings, then Your Claude (older layouts call it Memory), and click Manage next to Your memory. Everything Claude has saved about you is grouped by topic. Read it all, edit what’s wrong and delete what’s outdated. Memory is on by default on Free, Pro, Max and Ultra. (How memory works)

While you’re there, fill in the Instructions field. It applies to every chat you have.

Fill in your Project instructions

Projects work on every plan, and Free gets up to five. Make one for each workflow you repeat, click Set project instructions and spell out your role, rules, audience and files. Each project keeps its own memory, so your client work and your side project don’t blur together.

Tell Claude directly

Mid-chat, just say it. "Remember that I never want bullet points in emails." "Update your memory: I prefer replies under 400 words." "Forget that I mentioned [x]." Then check your memory page to confirm it stuck.

Import from ChatGPT or another AI

Go to Settings, then Your Claude, and click Start import. Claude gives you a prompt to paste into your old AI, which exports what it knows about you. Paste the answer back and click Add to memory. It works on every plan on web and desktop, and Anthropic calls it experimental. (Import your memory)

Or skip the typing. Paste this into a chat and fill in the brackets. You can also paste it into your Instructions field.

Full Memory Setup

I need you to remember the following about me. Save all of this to your memory so it applies to every future conversation we have. WHO I AM: - My name is [YOUR NAME] - My job title is [YOUR TITLE] at [YOUR COMPANY/BUSINESS] - I work in the [YOUR INDUSTRY] industry - I've been doing this for [X] years - My main responsibilities are: [LIST 3-5 KEY THINGS YOU DO] HOW I WORK: - My communication style is [direct / casual / formal / friendly but professional] - When I ask for writing, I want it to sound like [describe your voice, e.g. conversational? authoritative? warm? no-nonsense?] - I prefer [short and punchy / detailed and thorough / somewhere in between] responses - I hate when AI [uses corporate jargon / adds unnecessary caveats / gives vague answers / adds emojis / starts with "Great question!" (list your pet peeves)] - When I say "draft this," I mean ready to send, not a rough outline MY TOOLS: - I use [list the apps you use daily, e.g. Gmail, Slack, Notion, Google Docs, Salesforce, etc.] - My calendar is on [Google Calendar / Outlook / Apple Calendar] - I manage projects in [Notion / Asana / Monday / Trello / etc.] MY GOALS RIGHT NOW: - My #1 priority at work right now is: [WHAT YOU'RE FOCUSED ON] - A secondary goal is: [SOMETHING ELSE YOU'RE WORKING TOWARD] - By the end of this [month / quarter / year], I want to have: [SPECIFIC OUTCOME] WHAT I NEED FROM YOU: - Default to being [proactive / brief / thorough]. Don't wait for me to ask follow-up questions if the answer is obviously incomplete - If I ask you to write something, match my voice every time - Challenge my ideas when they have obvious flaws. Don't just agree with me. - If you're not sure about something, say so. Don't make things up. Save all of this to memory. Then show me a short list of exactly what you saved, so I can correct anything you got wrong. I should never have to explain any of it again.

Who This Is For

Layer 1 is enough for most people, and every piece of it works on the Free plan.

Layer 2 moves your memory into plain .md files on your computer that Claude reads every time you work. You own them, you can edit them, and you can hand the same files to any AI tool.

Instructions.md

Your rules. Who you are, what you do, how you want things done and what good output looks like. Include the line "Update Memory.md with my preferences over time." That’s what makes it a running log.

Memory.md

The brain. It starts mostly empty and grows as you work. Say "stop using em dashes" and Claude writes it down here. I split it into four sections, for preferences, corrections, patterns and decisions.

Context.md

The specifics for your business or project, like who your audience is, what you are working toward and who you work with. Make one general file or one per project.

Archive copies

Claude edits these files, and once in a while it overwrites something it shouldn’t. Once a week, copy the whole folder somewhere Claude can’t reach and label it with the date.

How To Use It

Make a folder called Claude Master Folder and put the files inside. In the Claude desktop app, connect that folder to your Cowork task or project. If your account has the new combined chat and Cowork experience, you can do it from any chat. This part needs a paid plan. You can also open Claude Code in the folder. (Cowork on your plan)

Build the files yourself, or connect the empty folder and paste this. Claude builds them for you.

Claude Builds Your Master Folder

You are setting up my AI memory system. Work in the "Claude Master Folder" I connected and build these four markdown files inside it: 1. Instructions.md: include these sections: ## Who I Am [My name, role, company, industry, years of experience] ## What I Do [My primary responsibilities and the type of work I do daily] ## Rules - Match my communication style: [direct / casual / formal / friendly but professional] - My writing voice is: [describe, e.g. conversational, authoritative, warm, no-nonsense, etc.] - Default response length: [short and punchy / detailed / somewhere in between] - Never use these words or phrases: [list your banned words, e.g. "leverage," "synergy," "it's worth noting," etc.] - Never start responses with: "Great question!" / "Absolutely!" / "Of course!" - Never add disclaimers like "As an AI..." unless I specifically ask - When I say "draft this" I mean ready to send, not bullet points or an outline - If I ask "what do you think?" give me your actual recommendation, not just pros and cons - If you're unsure about something, say so instead of guessing - Challenge my ideas when they have obvious flaws ## What Good Output Looks Like - Lead with the answer, then explain if needed - Use headers and bullets for anything longer than 3 paragraphs - Bold the most important point in each section - Write at an 8th grade reading level. Short sentences. Real words. Use contractions. - Keep paragraphs under 4 sentences ## Memory Rule Update Memory.md with my preferences, corrections, and patterns over time. Whenever I correct you, tell you I prefer something a certain way, or establish a pattern, log it in Memory.md so you never make the same mistake twice. 2. Memory.md: include these sections with placeholder examples so I know what to add: ## Preferences - [Example: Prefers bullet points over numbered lists when order doesn't matter] - [Example: Wants all emails to end with a clear next step] - [Example: Prefers casual tone in Slack, formal in client emails] ## Corrections - [Example: Stopped using em dashes after user asked; use commas or periods instead] - [Example: User corrected that their company name is capitalized as "TechFlow" not "Techflow"] ## Patterns - [Example: User typically asks for email drafts on Monday mornings] - [Example: User prefers to review outlines before full drafts on large projects] ## Decisions - [Example: Decided to use Notion instead of Asana for project tracking as of October 2026] - [Example: Changed email signature format; now includes phone number] ## Personal Context - [Example: User is based in EST timezone] - [Example: User's manager is Sarah, prefers formal updates to her] 3. Context.md: include these sections with a template format I can fill in: ## About This Project / Business [What the business does, who it serves, what makes it different] ## Audience [Who I'm talking to: clients, team, executives, customers, followers] [What they care about, how they prefer to communicate] ## Key People & Collaborators [Names, roles, and relationship context for the people I work with most] ## Active Projects & Priorities [What I'm working on right now, deadlines, milestones] ## Tools & Stack [The apps and tools I use daily and what each one is for] ## Important Background / History [Context that doesn't fit above but matters, like past decisions, company history, industry context] 4. Archive-Guide.md: a step-by-step guide explaining: - Why archiving matters (Claude can accidentally overwrite memory) - How to do it weekly (duplicate the Master Folder, rename it with the date like "Archive-2026-11-01", move it somewhere Claude can't access) - What to include in each archive (all .md files in their current state) - How to restore if something breaks (copy the archived file back into the Master Folder, overwriting the broken version) - Where to store backups (a separate folder on your desktop, a cloud drive, or anywhere outside the Master Folder) Make all four files clean, well-organized, and ready for me to fill in with my real information. Use markdown formatting throughout. Before you write anything, ask me up to 5 short questions to fill in the [brackets] above with my real information. Then write the four files, show me a one-line summary of each, and tell me which brackets I still need to fill in myself. Don't delete or overwrite any file already in the folder without asking me.

This one takes setup and upkeep, so it isn’t for everyone. But once it’s built, Claude works from everything you’ve written, read and worked on. Pick one of two options.

Claude + Notion, the easier option

Go to Customize (being renamed Plugins, so you may see that name), then Connectors, and connect Notion. Claude can now search and edit your pages and databases from any chat.

Make a Notion database called Memory Database. As you work, say "Send this to my Notion Memory Database." Over time you get a searchable knowledge base with Notion’s boards and filters on top, and you can export it as CSV for any other AI.

Claude + Obsidian, the most powerful option

Obsidian is a free notes app that saves everything as plain Markdown files in a folder called a vault. Because it’s just a folder, you connect it the same way as Layer 2, and Claude can read, write and link your whole knowledge base.

Download Obsidian and create a new vault.
Connect the vault folder in the Claude desktop app, or open Claude Code inside it.
Copy Andrej Karpathy’s LLM Wiki idea file. It lays out how an AI should build and maintain a linked wiki from your sources.
Paste it into the prompt below and send.
Feed it. Meeting notes, project briefs, client notes, articles and daily brain dumps. Each one makes the wiki smarter.

Build My Obsidian Second Brain

You are going to build and run my AI second brain inside my Obsidian vault, which is the folder I've connected. FIRST, READ THIS PATTERN: [paste the full text of Andrej Karpathy's "LLM Wiki" gist here] SET UP THE VAULT using that pattern: 1. Create three areas: a raw/ folder for my source files (you read these but never edit them), a wiki/ folder for the pages you write and maintain, and a schema file (CLAUDE.md) that records the rules for how this wiki works. 2. Create wiki/index.md (a catalog of every page, updated each time you add something) and wiki/log.md (a dated, append-only record of what you changed). 3. Add starter pages: wiki/me.md (who I am, how I work, my preferences), wiki/projects.md, and wiki/people.md. THEN, EVERY TIME I ADD A SOURCE (meeting notes, a project brief, an article, a client note, a daily brain dump): - Read it, write a short summary page, and update any related person, project or idea pages. - Link pages to each other with [[double brackets]] so Obsidian shows the connections. - Update index.md and add a line to log.md. WHEN I ASK A QUESTION: search the wiki first, answer with links to the pages you used, and offer to save a useful answer as a new page. ONCE A WEEK when I say "lint the wiki": list contradictions, outdated claims, orphan pages and gaps, and suggest fixes before changing anything. ABOUT ME TO START: [your name, role, current projects, key people, and how you like to work] RULES: - Never delete or overwrite a file in raw/. - Ask before deleting or merging any wiki page. - If a source contradicts an existing page, flag it instead of silently picking one. - Start now: set up the folders and files, then show me the structure and ask me for my first source.

Which One?

If you already live in Notion, start there. If you want the deepest setup and have an hour or two to build it, go with Obsidian.

How to Give ClaudePerfect Memory

How to Give Claude
Perfect Memory