Andrew Erickson · May 31, 2026

Build your LLM team wiki.

The shared brain your humans and your AIs are missing.

InventoryHero
1 / 51
About me

Andrew Erickson.

Andrew Erickson
  • 2010Graduated with an applied math & engineering degree.
  • 2015Started selling on Etsy and Amazon.
  • 2021First 7-figure Amazon exit.
  • 2023AI Grand Prize winner · 7-Figure Seller Summit.
  • 2026AI Grand Prize winner · Kevin King's Billion Dollar Seller Summit.
  • 100+Custom-designed products launched.
InventoryHero
2 / 51
Big news

I'm excited to announce… I just had my second successful exit!

InventoryHero
Andrew with his two kids in a Montessori classroom
3 / 51
What I'm building now

InventoryHero.ai + the community around it.

Now shipping
InventoryHero.ai
The agentic demand forecasting and supply chain tool for Amazon sellers. Amazon API in plain English: AI runs your entire supply chain.
Free community
AI for Ecom
WhatsApp group for sellers building with agents day-to-day.
Scan to join the free AI for Ecom community on WhatsApp
SCAN TO JOIN
Hosting
AI Agent Hackathon
Weekly builder series for AI-native ecommerce teams.
InventoryHero
4 / 51
The workshop

AI or Die.

AI is here.
Every company must use it.
Every employee must use it.
InventoryHero
5 / 51
Jensen Huang
Every company will be an AI company. Every employee will work alongside AI.
Jensen Huang
Founder and CEO, NVIDIA · paraphrased, GTC 2024
InventoryHero
6 / 51
The next 12 months

The best businesses will be 3 to 10 people, each with 3 to 10 AIs.

YOUR COMPANY CEO Marketing Ops Finance Product IH
InventoryHero
7 / 51
Marc Benioff
Every company will deploy thousands of agents alongside their employees.
Marc Benioff
CEO, Salesforce · paraphrased, Dreamforce 2024
InventoryHero
8 / 51
The question

So how do we build that?

One company. 10 humans. 50 agents.

How do they share a common knowledge base?

InventoryHero
9 / 51
Agenda

What you'll learn.

  • 01How AI memory actually works, in human terms.
  • 02Why every memory hack you've tried breaks at the team level.
  • 03The AI Employee Handbook framework.
  • 04Two ways to deploy it today, power-user and easy-button.
InventoryHero
10 / 51
Part 1 of 5
The Problem.

Your AIs have amnesia.

InventoryHero
11 / 51
The real problem

AI memory is broken, especially for teams.

Every chat
Starts from zero.
Even when set up
It lives on one laptop.
Your team
Can't see any of it.
InventoryHero
12 / 51
Quick Memory Sync

Don't wait until the talk ends.

Sign up for a free account at InventoryHero to set up your shared team memory.

Scan to sign up for InventoryHero
SCAN TO START
https://inventoryhero.ai/signup
01 · Setup time
Auth runs in the background.
02 · Live demo
Write your handbook, not mine.
03 · By minute 50
A working memory layer in 15 minutes.
InventoryHero
13 / 51
Module 1 of 5
In human terms

The anatomy of AI memory.

InventoryHero
14 / 51
Preview

The types of AI memory we'll cover.

  • 01Context Window
  • 02Context Window Rot
  • 03Vector Database
  • 04RAG (Retrieval-Augmented Generation)
  • 05LLM Wiki
  • 06Team Memory
InventoryHero
15 / 51
The framing

Think of your AI as an employee in their home office.

An AI employee at their home office, ready to work
InventoryHero
16 / 51
01 · Context

Context Window.

Sharp. Quick. Ephemeral.

What your AI is working on in this specific session.

  • 200K tokens = a long novel.
  • 1M tokens = the entire Harry Potter series.
InventoryHero
17 / 51
01 · Context

Context Window.

The desktop surface for your AI.
An AI employee's organized desk holding the documents they're working on right now
InventoryHero
18 / 51
01 · Context

Context Window best practices.

The organized desk: your AI's context window
  • <20%Sharp focus. Clear answers. Best work happens here.
  • ~50%AI starts forgetting earlier turns. Quality drops.
  • TipStart fresh chats for new tasks. Don't drag old context in.
  • TipBe selective. Paste only what's relevant, not "my whole Drive."
  • TipLong history? Summarize first, then continue.
InventoryHero
19 / 51
01 · Context

Context Rot.

When your desk gets buried.
A frustrated AI employee at a desk buried in papers, unable to find what they need

When your AI starts forgetting things inside the context window, that's context rot.

  • WhyLong inputs cause attention to spread thin across the prompt.
  • SignsEarlier instructions ignored. Repeated questions. Bad recall.
  • FixClear the desk. Summarize. Move durable knowledge out of chat.
InventoryHero
20 / 51
01 · Context

Context Rot, on a chart.

Output quality stays high until the window fills, then it falls off a cliff.
Chart: output quality stays around 95 below 40 percent context usage, drops gradually through the rotation point at 65 percent, then plunges through auto-compaction at 80 percent.
InventoryHero
21 / 51
02 · Vector

Vector Memory.

A library, with a table of contents the AI can read.
A library of books representing the AI's vector memory, with each book indexed and findable

Vector memory is like having a library of files.

  • ShelfEvery file your business has produced, indexed and stored.
  • VectorEach file gets a coordinate, like a book's title plus its table of contents.
  • ResultThe AI can find the right chapter without reading every book on the shelf.
InventoryHero
22 / 51
03 · RAG

RAG.

Retrieval-Augmented Generation.

The plumbing that lets your AI reach into your library and pull out only what it needs.

  • 1You ask a question.
  • 2The system searches the vector library for matching chunks.
  • 3Those chunks get pasted into the context window, just in time.
  • 4The AI answers using your data, not its training memory.
InventoryHero
23 / 51
03 · RAG

RAG is your librarian.

A librarian assistant pulling the right book off the vector library shelf for the AI

RAG is the assistant who runs to the library, grabs the exact right chapter, and walks it back to your desk.

  • SmartThe librarian uses the index. Doesn't read every book.
  • FastPages arrive on the desk only when the AI needs them.
  • CleanWhen the task is done, the pages go back. Desk stays clear.
InventoryHero
24 / 51
02-03 · Vector + RAG

Vector + RAG memory: tradeoffs.

Pros
  • Infinite scale. Searches billions of tokens beyond the context limit.
  • Controlled. Real-time data updates, instant security filtering.
  • Citations. Verifiable text matches reduce hallucinations.
Cons
  • Blind spots. Misses info if the search query or algorithm fails.
  • Lag. Multi-step pipeline adds response latency.
  • Complex. Requires building and tuning custom data infra.
  • Hard to gate. Once a file is indexed, it's hard to limit who can retrieve it.
InventoryHero
25 / 51
Andrej Karpathy
The hottest new programming language is English.
Andrej Karpathy
Founding member, OpenAI · ex-Director of AI, Tesla
InventoryHero
26 / 51
04 · LLM Wiki

LLM Wiki.

Karpathy's idea. Now everyone's running one.

A folder of ~50 markdown files. The AI reads them as truth.

No vectors. No embeddings. Just well-written notes anyone can author in an afternoon.

InventoryHero
27 / 51
04 · LLM Wiki

Plain text. Git-tracked. Free forever.

repo/
├── README.md           # the map
├── prompts/            # reusable templates
├── context/            # the Handbook lives here
│   ├── brand/
│   ├── products/
│   ├── marketing/
│   └── operations/
├── data/               # SP-API, reviews, sales
└── outputs/            # AI-written drafts
InventoryHero
28 / 51
04 · LLM Wiki

The curated company handbook.

A clean desk with a single leather-bound company handbook, sitting front and center

The reference handbook at the center of the desk. A single source of truth the AI reads from every session.

  • Shape~50 markdown files. Brand, products, ops, decisions, voice.
  • UseThe AI reads what it needs, when it needs it. No guessing, no hallucinations.
  • PowerWhen the handbook is sharp, every output is sharp.
InventoryHero
29 / 51
The InventoryHero handbook · 14 folders, ~71 files

Every facet of an FBA business. One shared brain.

Business core
  • 00-business
  • 01-strategy
  • 02-account-health
  • 11-legal-ip
Supply chain
  • 04-sourcing-suppliers
  • 05-inventory-planning
  • 06-logistics
  • 07-finance
Growth and ops
  • 03-catalog-listings
  • 08-advertising
  • 09-marketing-brand
  • 10-customer-service
  • 12-team-operations
  • 13-reference
InventoryHero
30 / 51
04 · LLM Wiki

LLM Wiki: tradeoffs.

Pros
  • Human-readable. You can audit and edit it. Your team can too.
  • No infra. No vector DB, no pipeline. Just files in a folder.
  • Portable. Works with any AI. Git-versioned. Copy-paste anywhere.
  • Sharp. Curated truth beats raw documents every time.
Cons
  • Manual. Someone has to write and maintain the handbook.
  • Bounded. Doesn't scale to millions of documents.
  • Drift. Goes stale unless you keep it current.
  • One brain. If it lives on one laptop, your team can't see it.
InventoryHero
31 / 51
05 · Team Memory

Team Memory.

The shared wiki for your humans and your AIs.

Take the LLM Wiki off your laptop. Put it in the cloud. Now every human and every AI on your team reads from the same handbook.

  • WhereIn the cloud, not on one laptop.
  • Who readsHumans and their AIs. ChatGPT, Claude, Gemini, your VAs, your suppliers.
  • Who writesHumans and AIs. Every interaction can update the handbook.
  • Who seesThe whole team. One source of truth, always in sync.
InventoryHero
32 / 51
05 · Team Memory

The shared handbook.

Multiple home offices, every desk holding the same shared handbook in sync

The same handbook on every desk. Updated by anyone. Read by everyone.

  • LiveUpdates sync to every team member instantly. No stale copies.
  • OpenHumans browse the wiki. Their AIs read the same files.
  • CompoundsEvery chat makes the handbook sharper. The team gets smarter together.
InventoryHero
33 / 51
 

Memory isn't a feature.
It's the difference between a VA on Day 1 and an ops manager on Year 3.

InventoryHero
34 / 51
Module 2 of 5
Why every solution you've tried fails your team

The hierarchy of memory tools.

InventoryHero
35 / 51
Memory hierarchy

Four tiers. Only one works for a team.

Frustrated executive at a messy desk overflowing with papers
Tier 1 · Ephemeral
A messy desk.
Cram everything in, and the AI starts missing details right in front of its face.
Executive with a bookshelf and a librarian assistant pulling a chapter
Tier 2 · Vector + RAG
A library + librarian.
A smart index finds the exact chapter, hands it over, and discards it when done.
Single executive at an organized desk with a private notebook
Tier 3 · Local vault
A great private notebook.
Beautiful for you. Invisible to your team and every AI not sitting at your desk.
Multiple home offices syncing the same handbook
Tier 4 · LLM wiki
One living handbook.
An AI librarian updates one 50-page digest. Every human and every AI reads it.
InventoryHero
36 / 51
Recap

The full memory stack, on one desk.

A complete home-office workspace with the desk, bookshelf, librarian, handbook, and shared cloud all labeled

Five layers. One workspace. Each layer does one job well.

  • Context WindowThe desk. Sharp, focused, ephemeral.
  • Vector DBThe library. Every file you've ever produced, indexed.
  • RAGThe librarian. Fetches the right chapter on demand.
  • LLM WikiThe reference handbook. Curated truth, always at hand.
  • Team MemoryThe cloud sync. Same handbook on every desk, every AI.
InventoryHero
37 / 51
What goes inside · 10,100 markdown files

Five categories. No 10,000 pages.

Who you are
Brand, mission, voice, avatar.
What you sell
Catalog, ASINs, USPs, pricing.
How you market
Listings, images, email, TikTok.
How you operate
Suppliers, freight, PPC rules.
Who's on the team
Humans and AI agents.
The rule
100 pages, not 10,000.
InventoryHero
38 / 51
Scaffolded, not schematic

Rigid databases die. Markdown adapts.

80% scaffolded
Folder structure + filenames are fixed.
The librarian always knows where things go.
20% fluid
Inside each file, AI writes freely.
Your brand's tone, decisions, and history evolve without migrations.
InventoryHero
39 / 51
Module 4 of 5
Karpathy-style · GitHub-backed · free

The power-user route.

InventoryHero
40 / 51
Two roads

Free and advanced. Or hosted and easy.

Road 1 · This module
Karpathy + GitHub
Free. Total control. Requires the terminal.
Road 2 · Module 5
Hosted MCP
5-minute setup. Mobile-friendly. Non-technical team.
InventoryHero
41 / 51
Why GitHub works as memory

Shockingly good. And free.

Version history
Never lose context.
Pull requests
A paper trail for every change.
Collaborators
Instant team access.
Tool support
Claude Code, Cursor, Codex, Gemini.
Cost
Free for private repos.
Sync
Push here. Pull there. Done.
InventoryHero
42 / 51
Live demo

GitHub as memory. In seven moves.

  • 01Create a private repo: my-brand-handbook.
  • 02Drop in the Karpathy folder structure.
  • 03Open it in Claude Code.
  • 04"Write 5 TikTok Shop hooks for ASIN B0XXX."
  • 05Review · commit · push.
  • 06On laptop two: git pull. Memory is there.
  • 07Add your VA as a collaborator. Same brain.
InventoryHero
43 / 51
Is this path for you?

It depends on who's on your team.

Yes, if
  • You're comfortable in a terminal.
  • You use Cursor or Claude Code daily.
  • You want $0 in monthly cost.
  • You sell products and ship code.
No, if
  • You've never run git push.
  • Your team is mostly VAs and creatives.
  • You want this working in 5 minutes.
  • You want it on your phone.
InventoryHero
44 / 51
Module 5 of 5
Same framework · zero setup · five minutes

The easy button.

InventoryHero
45 / 51
The architecture

Two tools. One librarian. One database.

Claude · Gemini · Cursor · ChatGPT
↓ MCP read / write
InventoryHero MCP
read_memories
write_memories
↓
Sonnet 4.6 librarian
↓
Postgres · tenant-isolated · full revision history
InventoryHero
46 / 51
Onboarding interview

You don't fill 71 blank files. It interviews you.

  • 01Your AI runs a Socratic interview about your brand.
  • 02You answer in plain English while you work.
  • 03The librarian files each answer into the right docs.
  • 04Upload PDFs and supplier sheets, it absorbs them.
  • 05Watch /handbook fill up over a few days.
InventoryHero
47 / 51
Audience reveal

If you signed up at slide 17, you're already live.

  • ·Signing up now? Set up by slide 60.
  • ·Signing up at Q&A? Set up by the time you stand.
  • ·8-figure sellers, you'll break us. We'll wear it like a badge.
Scan to sign up for InventoryHero
SCAN TO START
https://inventoryhero.ai/signup
InventoryHero
48 / 51
What we covered

Four ideas. One outcome.

01
AI memory ≈ human memory.
Working vs. long-term.
02
Local vaults are the PDFs of AI memory.
03
The AI Employee Handbook framework.
Scaffolded, progressive, ~100 files.
04 · The outcome
Every AI in your business reads from the same brain.
InventoryHero
49 / 51
Build this with me · 100% free

AI Hackathon. Tuesday, June 2.

9 AM PT · 12 PM ET · 5 PM London

  • ·Spin up your Employee Handbook.
  • ·Wire up the GitHub path (Road 1).
  • ·Install the InventoryHero MCP (Road 2).
  • ·Live Q&A with the dev team.
inventoryhero.ai/hackathon
Bring a laptop and a coffee.
InventoryHero
50 / 51
Over to you

Q & A.

InventoryHero
51 / 51