The MEMORY.md File: How Claude Code Remembers You Between Sessions

Claude Code for Founders · Episode 3: Memory. 47 seconds. The MEMORY.md file that persists across sessions. Watch on YouTube.

The amnesia problem. Every new chat starts from zero.

Open ChatGPT. Tell it your brand voice. Tell it what you are working on this week. Tell it the correction you made yesterday. Get a draft. Close the tab. Open a new chat tomorrow. The AI has no idea who you are.

This is the amnesia problem, and it taxes every founder using AI to run a content operation. You spend the first ten minutes of every session re-pasting context the AI already had last time. The dashboard you described, the brand voice you locked in, the typo you corrected three sessions ago, the project you have been building for a month: all of it gets pasted again. And again. And again.

The cost is not just the ten minutes. The cost is that the AI you work with is a brilliant stranger every single conversation. It cannot learn what you want, because it cannot remember what you told it.

We covered the broader version of this problem in Brand as Code (Episode 1) and the file-level fix for brand voice in The File (Episode 2). This post is the missing piece: how Claude carries that file forward, plus everything else worth remembering, between sessions.

What a MEMORY.md file actually is.

A MEMORY.md is a plain markdown file that Claude Code maintains for itself on your machine. You can open it. You can read it. You can edit it. You can commit it to git. Claude reads it at the start of every session and proposes additions to it at the end of relevant ones.

It is not a vector database. It is not opaque chat history. It is a single, human-readable file of facts the AI decided were worth remembering about you, your projects, and the decisions you have made.

Where CLAUDE.md (the brand voice file from Episode 2) is the spec that does not change, MEMORY.md is the diary that does. The voice file says how to write. The memory file says who you are right now and what you are in the middle of.

What it remembers. Three categories.

1 · Preferences

How you like to work. Your communication style. The tools you reach for. The cadence you ship on. These are facts about you that should never need to be re-explained, no matter how many sessions go by. Once they are in MEMORY.md, every future session loads them automatically.

2 · Projects

What you are building right now. The launch this week. The campaign running on the side. The PR you are reviewing. These change weekly, but the AI always knows what you are in the middle of, so it can connect today's request to last week's context without you having to set the scene.

3 · Corrections

The things you told the AI never to do again. The phrase that does not fit your brand. The factual mistake it made yesterday. The shortcut that broke production last week. Each correction gets written down once, kept once, and never re-explained.

The third category is the one that compounds fastest. Most AI tools treat your corrections as forgotten the moment the conversation ends. Claude treats them as load-bearing rules in the file every future session reads.

How memory gets written.

You do not write to MEMORY.md by hand for routine facts. Claude proposes additions during sessions when it notices something worth keeping. You see the proposal as it happens and either accept or skip. Accepted facts get appended to the file. The memory grows organically as the AI learns who you are.

The crucial part: you have full editorial control. If something looks wrong, you edit MEMORY.md directly with a text editor. Claude re-reads on the next session, and your correction is permanent. Nothing is hidden in a database you cannot inspect. The memory is yours.

This matters because content operations live or die on the corrections. The whole point of accumulated memory is that "do not call our customers users" gets written down once, not re-typed every time the AI forgets. A file you control gives you that. An invisible memory layer does not.

How memory gets recalled.

At the start of every new session, Claude reads the MEMORY.md file before producing anything. The facts get loaded into context the same way the brand voice file does. By the time you type your first prompt, the AI already knows your preferences, your active projects, and the corrections you have made.

You stop starting from zero. You start from where you left off. Multiply that across a year of sessions and the gap between Claude and a stateless chatbot becomes unbridgeable. Claude knows you. The other one is still meeting you for the first time.

Why this beats vector stores and chat history.

The two common ways AI tools try to "remember" you are vector databases and chat history. Both are weaker than a file.

Vector databases store semantic chunks and retrieve them on similarity. They can find what you said about your launch last month if you ask a related question. What they cannot do reliably is enforce a hard rule like "never use the word transformative." That is not a similarity search, it is a constraint. Rules belong in a file, not a database.

Chat history is even weaker. It is opaque. You cannot read what the AI considers important about you. You cannot edit it. You cannot version-control it. You cannot share it with a teammate. When the system gets it wrong, you have no recourse other than starting a new conversation.

A markdown file solves all three problems at once. You see what gets remembered. You edit what is wrong. You commit it to git so it travels with the project and the history. Memory becomes infrastructure, not magic.

A copyable MEMORY.md starter template.

Here is a minimal MEMORY.md block you can drop into a Claude Code memory directory and start growing. The shape is what matters. The contents will accumulate over time.

# Your Name · working memory

## Preferences
- writes in: direct, no-nonsense, executive
- ships on: Mon / Wed / Fri
- formats: one thesis per piece, lead with the claim

## Projects
- whystrohm.com · launch week
- NVUS · Meta ad sprint

## Corrections
- never pad the intro
- never use the word transformative
- ship before optimizing

## Recalled facts (auto-appended)
# Claude appends learned facts here. Edit freely.

The last section is the one Claude proposes additions to. Everything above it you maintain by hand. The file grows over time as the AI learns your shape, and you can audit it weekly the same way you audit any other source-controlled file in your project.

How to start.

1. Run a Claude Code session in your project. Claude creates the memory directory automatically the first time it has something worth remembering.
2. Drop the template above into the MEMORY.md file in that directory.
3. Edit the three sections (Preferences / Projects / Corrections) to match your reality. Two minutes of work the first time, zero from then on.
4. Work with Claude normally. When it proposes a memory addition, accept the ones worth keeping. Reject the rest. Audit the file weekly.

This is Episode 3 of Claude Code for Founders. Episode 1 covers why brand voice should live as code. Episode 2 covers the brand voice file. Episode 4 covers skills, the named commands that turn a workflow into a single line of input.