Three files. Twenty minutes. Claude never forgets your business again — niche, tone, offer, and history load automatically at every session start.
This guide is reviewed for clarity, service accuracy, and AI-search readability. The next quarterly content review is tracked internally before unsupported metrics or client proof are added.
The Problem You Keep Running Into
Every new Claude chat starts blank. No memory of your business. No memory of your tone. No memory of the decision you made last week. You re-explain everything, every time.
This is not a bug. Claude is stateless by design. But there is a fix that takes about 20 minutes to set up once — and then Claude never asks you to catch it up again.
Three files. That is the whole system.
The Architecture
prompt~/.claude/ ├── CLAUDE.md ← loads automatically at every session start ├── MEMORY.md ← lightweight index of everything in facts/ └── facts/ ├── user_profile.md ├── project_context.md ├── feedback_preferences.md └── reference_links.md
CLAUDE.md is the briefing document. Claude Code loads it automatically every time you open a session — no prompt needed. This is where your identity, niche, tone, offer, and workflow rules live.
MEMORY.md is the index. It stays short (under 200 lines) and lists every fact file with a one-line description of what is inside. Claude reads this to know where to look without loading everything at once.
facts/ is the library. One file per topic. Claude pulls individual files only when they are relevant — it does not dump the entire folder into every session.
How Claude Reads These Files
Claude Code (terminal or IDE): CLAUDE.md at ~/.claude/CLAUDE.md loads automatically at session start. No action required. Files you reference inside CLAUDE.md using @path syntax also get pulled in. This is Claude Code's native memory system.
Claude.ai Projects (web interface): Upload CLAUDE.md and MEMORY.md as Project Knowledge files inside a Project. Claude reads them at the start of every conversation in that project. For facts/ files, either paste their contents into a single combined Knowledge file or reference them inline inside CLAUDE.md.
The key difference: Claude Code can read and write files at real paths. Claude.ai Projects stores knowledge you upload — Claude can output updated content for you to paste in, but it cannot write files directly.
Step-by-Step Setup
Step 1: Create the folder structure
If you are using Claude Code, run this once in your terminal:
bashmkdir -p ~/.claude/facts touch ~/.claude/CLAUDE.md touch ~/.claude/MEMORY.md touch ~/.claude/facts/user_profile.md touch ~/.claude/facts/project_context.md touch ~/.claude/facts/feedback_preferences.md touch ~/.claude/facts/reference_links.md
Step 2: Fill in CLAUDE.md
This is the file Claude reads first, every session. The more specific you make it, the better Claude performs. Copy this starter and replace everything in brackets:
markdown# Identity I'm [YOUR NAME], founder of [BUSINESS NAME]. We [one sentence: what you do and who you serve]. # My Audience [2-3 sentences: who your ideal customer is, what they struggle with, what they want] # My Offer - [Product or Service 1] — [what it does in one line] - [Product or Service 2] — [what it does in one line] # Tone and Voice [Describe how you write. Example: Direct and practical. No fluff. Short sentences. Sound like a founder talking to a peer, not a marketer talking to a prospect.] # Content Pillars - [Pillar 1] - [Pillar 2] - [Pillar 3] # Workflow Preferences - [e.g., Give me 3 headline options before writing the full piece] - [e.g., Ask what I'm optimizing for before suggesting a strategy] - [e.g., Use bullet points over long paragraphs] # Memory System My memory files are in ~/.claude/facts/. The index is at ~/.claude/MEMORY.md. When you learn something new about my business, preferences, or decisions that should persist across sessions: 1. Save it to the right file in ~/.claude/facts/ using the frontmatter format 2. Add a one-line pointer entry in ~/.claude/MEMORY.md Save immediately when I say any of these: - "remember this" - "save this" - "note that" - "add this to memory" - "don't forget this" @~/.claude/MEMORY.md
The @~/.claude/MEMORY.md line at the bottom tells Claude to pull in the index file on every load.
Step 3: Set up MEMORY.md
MEMORY.md is a living index. It starts short and grows as Claude updates it. Each entry is one line pointing to a facts/ file:
markdown# Memory Index ## User - [My Profile](facts/user_profile.md) — business context, background, and goals ## Project - [Current Projects](facts/project_context.md) — active work, deadlines, decisions ## Feedback - [Preferences](facts/feedback_preferences.md) — how I like Claude to work with me ## Reference - [Key Links](facts/reference_links.md) — tools, platforms, and resources I use regularly
Step 4: Set up the facts/ files
Every file in facts/ uses the same frontmatter format. The name and description fields tell Claude when to pull the file. Be specific in the description — vague descriptions mean Claude will not know when the file is relevant.
facts/user_profile.md
markdown--- name: user-profile description: Who I am, my business, niche, and founding context metadata: type: user --- [Your name] is the founder of [Business Name]. **Business:** [What you do in one sentence] **Niche:** [Your target audience and what they care about] **Background:** [Relevant experience — keep it to what shapes how you work] **Goals:** [What you're building toward in the next 6-12 months]
facts/project_context.md
markdown--- name: project-context description: Active projects, current priorities, and recent decisions metadata: type: project --- **Active projects:** - [Project 1] — [what it is, current stage, why it matters] - [Project 2] — [what it is, current stage, why it matters] **Current priority:** [The one thing that matters most right now] **Recent decisions:** - [Decision] — **Why:** [reason] **How to apply:** [what this means going forward]
facts/feedback_preferences.md
markdown--- name: feedback-preferences description: How I like Claude to behave — confirmed approaches and things to avoid metadata: type: feedback --- **Do:** - [e.g., Give 3 options for hooks before writing the full caption] - [e.g., Start drafts immediately — don't ask for permission to write] **Don't:** - [e.g., Add lengthy disclaimers before answering] - [e.g., Suggest I "consult a professional" for business strategy questions] **Confirmed approaches:** - [e.g., Short bullet-point summaries over prose — confirmed preference after 10+ sessions]
facts/reference_links.md
markdown--- name: reference-links description: Tools, platforms, and resources I use regularly metadata: type: reference --- **Tools:** - [Tool name] — [what I use it for] **Platforms:** - [Platform] — [context, e.g., "my main content channel"] **Resources:** - [Resource name] — [why it matters to my work]
The Self-Update Loop
This is the part that changes everything. Once your files are set up, Claude updates its own memory as you work. You do not maintain a spreadsheet or manage a prompt document.
The loop works like this:
- You work with Claude on something real — writing copy, making a decision, testing a new workflow
- Something worth keeping comes up — a confirmed preference, a project update, a tool you just started using
- You say "remember this" (or any of the trigger phrases in your CLAUDE.md)
- Claude writes the fact to the right file in facts/ using the correct frontmatter format
- Claude updates MEMORY.md with a pointer if the fact is in a new file
- Next session, that fact is already loaded
The sessions compound. Claude gets sharper every week because the context is cumulative, not reset.
Real Examples
Here is what a filled-in CLAUDE.md looks like in practice — this level of specificity is what makes Claude actually useful without hand-holding:
markdown# Identity I'm Alex, founder of Hypd Agency. We help solo founders build AI-powered content systems that create and publish content without a full team. # My Audience Early-stage solo founders running service or info-product businesses. They know AI tools exist but don't have time to figure out how to chain them together. They want leverage, not more things to learn. # My Offer - AI Content Systems — done-for-you automation builds for Instagram + LinkedIn - Hypd Vault — monthly membership with templates and walkthroughs # Tone and Voice Direct. No fluff. Short sentences. Sound like I'm texting a smart friend who runs their own business — not a marketer writing a LinkedIn post. Concrete over abstract. Never use the word "leverage" unless I say it first. # Content Pillars - AI automation for founders - Building in public - Systems over hustle
Here is what a MEMORY.md index looks like after a few weeks of sessions:
markdown# Memory Index ## User - [My Profile](facts/user_profile.md) — business context, background, and goals ## Project - [Active Projects](facts/project_context.md) — current work and key decisions - [Launch Timeline](facts/q3_launch.md) — Hypd Vault launch plan and constraints ## Feedback - [Preferences](facts/feedback_preferences.md) — workflow and output preferences - [Content Voice](facts/content_voice.md) — confirmed tone patterns from reviewed drafts ## Reference - [Key Tools](facts/reference_links.md) — tools and platforms I use regularly
Here is what a fact Claude writes after you say "remember this — always lead Instagram captions with a pattern interrupt, not a question":
markdown--- name: instagram-caption-structure description: Confirmed Instagram caption opening — pattern interrupt leads, not questions metadata: type: feedback --- Always lead Instagram captions with a pattern interrupt — a bold statement, counterintuitive claim, or unexpected fact. Never open with a question. **Why:** Confirmed preference after reviewing 10+ drafts. Question-led captions underperform on reach. **How to apply:** Any Instagram caption draft starts with a statement that stops the scroll, not a question that invites the reader to scroll past.
Maintenance: When to Prune
The system stays fast when files stay lean. Check every 4-6 weeks:
- Stale projects — remove or archive entries in
project_context.mdfor completed work - Outdated preferences — if your workflow changed, update
feedback_preferences.mdso old rules stop overriding new behavior - Conflicts — if a fact file says one thing and the current reality says another, trust what you observe now and update the file; Claude verifies current state before acting on memory
- Duplicate entries — if MEMORY.md has two entries covering the same territory, consolidate them into one file
The signal that it is time to prune: Claude gives you advice based on something you stopped doing months ago. Find the stale fact file. Update it. Done.
Claude.ai Projects Setup (No Terminal Required)
If you are not using Claude Code, here is the web-based equivalent:
- Go to claude.ai → Projects → New Project
- Name the project after your business
- Open Project Knowledge and add:
- Your CLAUDE.md content
- Your MEMORY.md content
- Your facts/ file contents (paste each one under a clear header, or as separate knowledge files)
- Every new conversation inside that project loads all of it automatically
The self-update loop still works here — say "remember this" and Claude outputs the updated file content for you to paste into Project Knowledge. It cannot write files directly in the web interface, but the result is the same.
FAQ
Does this only work with Claude Code, or does it work in the regular Claude web app too?
Both work, but Claude Code is significantly more powerful for this system. Claude Code reads and writes files directly at ~/.claude/ — CLAUDE.md auto-loads every session and Claude can update your facts/ files in real time when you say "remember this." The web app (Claude.ai Projects) stores the same content as Project Knowledge files you upload, but you paste updates in manually. If you use Claude regularly for business work, Claude Code is worth installing.
What if Claude starts saving things to memory that I did not ask it to save?
Add this line to your CLAUDE.md inside the Memory System section: Only save to memory when I explicitly say one of the trigger phrases above. Do not save autonomously. This makes memory saves fully opt-in. The default system described here is more proactive — adjust it to match how you want to work.
How do I keep different projects from bleeding into each other?
Your global ~/.claude/CLAUDE.md handles identity and universal preferences — things that are always true about you. For project-specific context, add a CLAUDE.md inside each project folder (e.g., ~/projects/client-work/CLAUDE.md). Claude Code loads both files and merges them, with the project-level file taking precedence on any conflicts. Your niche, tone, and offer stay global. Project-specific deadlines, decisions, and constraints stay local.
This is the memory system Hypd uses across every AI workflow we build. The first session feels the same as before. By session five, you stop noticing the setup because Claude already knows what you are going to ask. Build it once. Let it compound.