The Product SystemThe Vault
AI Memory for Product Managers: 5 Levels So You Never Re-Explain Your Product, the infographic in this PDF

AI

AI Memory for Product Managers: 5 Levels So You Never Re-Explain Your Product

Five levels of AI memory for product managers, from Claude Code rule files to a product wiki, RAG search, evals and one agent, so you never re-paste context.

Want this as a PDF?

You'll also get The Product System, my weekly note on leading product in the AI era. Unsubscribe anytime, and I'll never sell your email. How I handle your details

1

The full infographic, as a PDF

The exact one from the post, sized to print or keep within reach.

2

A copy in your inbox

The download link lands in your email too, so it's there when you need it. No follow-up campaign.

3

The Product System, weekly

My note on leading product in the AI era. One idea a week that compounds. Unsubscribe anytime.

I explained my product once. My AI recalls it at 99.5%.

For months I pasted the same context from Notion into a chat window. Every single chat. The strategy, the customer, the decision we made last month and why. Then the window closed and the AI forgot all of it.

A product manager's value sits in context: why this customer and not that one, what you said no to, which number you steer by. If your AI can't hold that context, you end up as its copy-paste service, and the answers stay generic because the input is generic.

So I built a system instead. It has five levels. Each one fixes a problem you probably already feel every week, and you don't need all five on day one.

Level 1: Claude Code remembers you

A chat window forgets you when it closes. Claude Code reads your files before you type a word.

Start with one folder holding three files: CLAUDE.md, AGENTS.md and SOUL.md. Between them they say who you are, how you work, and your never-do list. Then type /memory to check that auto memory is on, so it saves what it learns as you go.

The test is simple. Open a new session and ask what it knows about you. If it gets you wrong, a line is missing from SOUL.md. Add the line and ask again. You are fixing the source, not arguing with the output.

Level 2: Your wiki remembers your product

At level 1 your AI knows you. It still can't tell you why you picked this customer over that one.

Make one folder for each stage of the product manager's job, plus an INDEX.md that the AI reads first. Then write PRODUCT.md before anything else: your strategy, your customer, and the one number you steer by. Rough is fine. A rough page it can read beats a polished one in your head.

From then on, file every interview and every decision the day it happens, with the why. The why is the part that matters most, because it is what lets the AI answer "why did we pick X?" months later. The idea comes from Andrej Karpathy's LLM Wiki: a folder of plain pages the AI reads, instead of context scattered across old chats.

Level 3: RAG search remembers everything

Past a hundred or so files, the index stops being enough. You can't expect the AI to open every page to find one line.

Search by meaning fixes that. It finds the note even when your question uses different words from the note itself. Ask Claude to set up the memory-search skill. By default it runs on your own machine, free and private. If you want it in the cloud, point it at Pinecone or Supabase instead.

Schedule a nightly re-index so new notes are found by morning. Then ask the questions that used to cost you an afternoon of digging: "Have we heard this complaint before?" or "What did we decide in March?"

Level 4: Evals make sure it remembers right

Finding a note isn't the same as the note being true. A memory system that answers confidently from a stale page is worse than one that says it doesn't know.

Write 20 questions you already know the answers to. Your customer. Your last big call. What you are not doing this quarter. Then run the memory-evals skill every week. It answers from the folder only and logs the score.

Aim for 99.5% over time. That is the recall score my own system reached. A wrong answer shows you which file to fix. Fix the file, never the answer. If you patch the answer, the same gap will show up in the next question you didn't think to test.

Level 5: One agent remembers everywhere

A memory only you can reach still runs through you. If your team has to ask you so that you can ask the AI, every answer still waits on your calendar.

For you, use Claude Code with Remote Control, so you can ask from your phone. For your team, set up a Claude or ChatGPT Project loaded with only your shareable files. Or drop Claude into Slack on a Team or Enterprise plan.

Whichever you pick, make it cite the file for every answer. Then anyone can check it, and a wrong answer points straight at the page that needs fixing.

Build the level you need next

Read the five levels together and the pattern is clear. Each level answers the question the one before it leaves open. Level 1 knows you but not your product. Level 2 knows your product but can't search it at scale. Level 3 searches it but can't tell you if it is right. Level 4 checks it is right but only for you. Level 5 hands it to everyone.

You don't need all five on day one. Start with the level that fixes the problem you feel most this week. Build them once, and you never re-explain your product again.

There is a craft point under the tooling too. Writing PRODUCT.md, filing decisions with the why, and writing 20 questions with known answers are product management habits first. The AI just makes it obvious when they are missing. If you want the wider argument for building a system around your role rather than buying the next model, see /vault/best-ai-system-cannot-be-bought.

Questions people ask

How do I give AI memory of my product as a product manager?
Put your context in files the AI reads every session instead of pasting it into chats. Start with rule files about you (CLAUDE.md, AGENTS.md, SOUL.md) in Claude Code, then add a product wiki with an INDEX.md and a PRODUCT.md covering strategy, customer and the one number you steer by. File decisions and interviews the day they happen, with the reasoning.
What should go in a PRODUCT.md file?
Your strategy, your customer, and the one number you steer by. It does not need to be polished. A rough version the AI can read is more useful than a perfect version that lives only in your head, and you can improve it as the AI's answers show you what is missing.
When does a product manager need RAG search?
When the wiki grows past a hundred or so files and an index is no longer enough for the AI to find the right page. Search by meaning finds a note even when your question uses different words. It can run on your own machine for free, or in the cloud through a service such as Pinecone or Supabase, with a nightly re-index.
How do you test whether your AI remembers your product correctly?
Write 20 questions you already know the answers to, such as who your customer is or what you are not doing this quarter. Have the AI answer from your files only, every week, and log the score. When it gets one wrong, fix the file that caused the miss, never the answer itself.
How can my team use my AI memory without asking me?
Give them one agent that reads your shareable files, such as a Claude or ChatGPT Project, or Claude in Slack on a Team or Enterprise plan. Make it cite the source file for every answer so anyone can check it. For yourself, Claude Code with Remote Control lets you ask from your phone.
Do I need all five levels at once?
No. Each level fixes a specific problem, so start with the one you feel most often. Level 1 alone stops you pasting the same context into every chat, and each later level builds on the one before.