Build Your
LLM Team Wiki

The shared brain your humans and your AIs are missing.

Andrew Erickson

Founder, InventoryHero.ai

May 31, 2026 · Austin, TX

About Me · Part 1

Andrew Erickson

Quick credentials, so you know I'm not a guru.

Started on Etsy/Amazon in 2015
Founder of three 7-figure brands on Amazon
Studied Applied Math & Electrical Engineering
Engineer turned Entrepreneur
Host of Titan Weekly AI Mastermind
Kevin King's BDSS 2026 Grand Prize Winner
Designed & launched 100+ physical products
7-figure exit in 2021
and…

About Me · Part 2

What I'm building now

Active Build

InventoryHero.ai

Agentic inventory management for Amazon sellers.

Talks to your SP-API in plain English
Detects stockouts, overstock, reorder risk before you do
The memory layer behind it is what we're building today

▶Host of the OpenClaw Hackathon series
▶Host of the AI for Ecom WhatsApp chat group
▶2nd successful exit just closed, more on that in a sec…

The launches I'm proudest of

Aliza & Arlo

Different launch, they're my kids.

Aliza

April 2020

Arlo

September 2022

Also…

"I just had my second successful exit.
Which means…
I'm officially a jobless bum.
So I'm filling my time by giving away the playbook."

Let's talk business.

My goal in the next 60 minutes

Blow your mind.

Pay attention to this one

How well you understand the next 60 minutes will determine whether your business is running 10 AIs in 12 months…
or being run over by them.

The Amazon sellers who own their context will own their category.

What you're going to learn

The next 60 minutes, in four pieces.

How AI memory actually works.
In human terms, no math, no jargon.
Why every "memory hack" you've tried is broken at the team level.
Local vaults are the PDFs of AI memory.
The AI Employee Handbook framework.
The named system you'll leave with today.
Two ways to deploy it today. ○ Power-user route, Karpathy + GitHub (free, advanced) ○ Easy button, hosted MCP (5 min, non-technical)

The hottest new programming language is English.

Andrej Karpathy founding member, OpenAI · ex-Director of AI, Tesla

Part 1 · The Problem

Your AIs have amnesia.

Every chat starts from zero. Every teammate works from a different memory. It's costing you more than you think.

The Problem · The real one

AI memory is broken, especially for Amazon teams.

Every new chat starts from zero.
Even when you DO set up memory, it lives on one laptop.
Your VAs, your designer, your PPC manager, your supplier, your 10 background agents, none of them can see it.
You re-explain your brand, your avatar, your ASINs, your TikTok angle, your email voice, to every AI, every day, forever.

Two Failure Modes

You can't out-prompt this.
You have to out-structure it.

− Too Little Context

"Update my listing copy."

Result: generic hallucination that could be for any seller in any niche.

+ Too Much Context

"Just read my entire Google Drive."

Result: noise, dilution, confusion.

The Team Problem

You're conducting a symphony where everyone has different sheet music.

You

Local Obsidian Vault

↓

Genius context that nobody else can read.

Your ops manager doesn't have it.
Your supplier in Shenzhen doesn't have it.
Your TikTok Shop VA doesn't have it.
Your PPC bot doesn't have it.
The 10 Claudes running your product research don't have it.

You don't need a smarter AI.
You need a shared brain.

AI without your data is just a parlor trick.

Aaron Levie co-founder & CEO, Box

Before we go any further

Sign up now, your account syncs while I talk.

Later in this talk I'll do a live demo using YOUR Amazon SP-API data.

If you scan this QR code right now, your account will sync in the background while I'm talking. By minute 45, your store will be live, your catalog loaded, and you'll be building alongside me. For free.

No credit card Beta access ~3 min to authorize SP-API

Scan to sign up

inventoryhero.ai/signup

Why now, not later

Three reasons to sign up while you listen.

SP-API sync takes ~20 minutes. Start it now and it's done before we need it.
Watch the AI Employee Handbook auto-populate from YOUR real data, not a demo account.
By minute 50, you'll have a working memory layer you can use Monday morning.

If you wait until the end of the talk, you'll have to go home and do this on your own.
Don't.

Module 1 · 7 minutes

The Anatomy of
AI Memory.

In human terms. No vector math. No jargon.

The Anatomy

Don't think "vector embeddings."
Think about how YOU remember.

Human Brain	AI Equivalent	Example
Working Memory	Context Window	The chat you have open right now
Long-Term Memory	Vector DB / RAG retrieval	Pinecone, Chroma, your wiki
Notebook on your desk	Markdown files	`brand_guidelines.md`
Calling a friend	MCP tool call	"Hey SP-API, what's stock on B0XYZ?"

Working Memory

The Context Window = Working Memory

▶Short-term, in-the-moment focus.
▶Incredibly sharp, Claude can hold massive content in one conversation.
▶Completely ephemeral, gone the moment the chat ends.
▶This is why "memory" doesn't survive a new conversation. It was never actually saved.

The chat ends.
The memory dies.
Every. Single. Time.

Long-Term Memory

Vector Embeddings + Retrieval

▶Your knowledge becomes coordinates in N-dimensional space.
▶Filing cabinet organized by meaning, not alphabet.
▶When you ask a question, the AI fetches relevant chunks…
▶…and pulls them INTO working memory, on demand.

This is how you give an AI a past.
This is how you stop starting from zero.

Memory isn't a feature.

It's the difference between
a VA on Day 1
and an ops manager on Year 3.

The bottleneck isn't model capability anymore, it's context.

Aravind Srinivas co-founder & CEO, Perplexity

(paraphrased from his 2025 interview appearances)

Module 2 · 8 minutes

How Context
Actually Works.

What fills it up. What gets forgotten. The numbers nobody tells you.

Context = the AI's RAM

Every chat has a token budget.

1 token ≈ 4 characters ≈ ¾ of a word. 200K tokens ≈ a 500-page book. Sounds infinite. It isn't.

Claude Opus 4.7advertised 1M · sharp recall ~200K

20%

GPT-5advertised 400K · sharp recall ~100K

25%

Gemini 2.5 Proadvertised 2M · sharp recall ~200K

10%

Sharp recall zone Fuzzy / degraded recall

"1M context" is a marketing number. Effective context is what your AI actually uses well.

Lost in the Middle

AI recall is a U-curve, not a flat line.

Position of fact in prompt →

Sharp at the start. First ~10K tokens get full attention.
Sharp at the end. Last ~5K tokens (recency).
Hazy in the middle. Recall drops 20 to 40% on facts buried mid-prompt.
▶Documented in Liu et al. 2023, "Lost in the Middle" and reproduced on every frontier model since.

Same prompt. Same model. Same question. Move the answer to the middle, and it gets it wrong.

What's eating your context

Your data is the thin slice. Most of the budget is overhead.

A typical agentic chat with 3 MCPs connected, before you've even typed your question:

Sys

Tool definitions

Conversation history

Retrieved docs / RAG

System prompt
1K to 10K

Tool defs (MCP)
5K to 30K each

Chat history
grows linearly

Retrieved docs
10K to 80K

Your question
~200 tokens

By turn 10, the model is reading 90K tokens of overhead to answer 200 tokens of you.

The implication

Never dump. Always curate.

What most teams do

Dump and pray.

"Just read my entire Google Drive."
Pile 14 PDFs into one chat.
Re-paste the same SOPs every prompt.

Result: the answer they cared about lands in the dead-zone.

What the Handbook does

Map + pull.

One small handbook_overview.md in the sharp zone.
AI pulls only the 1 to 3 files it needs.
Sharp recall on the part that actually matters.

Result: the model thinks with your business, not against it.

The Handbook is a context-budget strategy, not a wiki.

Module 3 · 7 minutes

The Hierarchy of
Memory Tools.

Why every solution you've tried has failed your team.

The Memory Hierarchy

Four tiers. Three of them fail your team.

Type	Implementation	Good for	Where it breaks
Ephemeral	Copy-paste · system prompts	One-off chats	Manual. Every. Time.
Siloed Static	Obsidian · LogSeq on laptop	Solo founders	Can't share across team
Vector Space	Pinecone · Chroma · raw RAG	Engineers	Humans can't read it
Dynamic Shared	Team MCP + cloud markdown	Real businesses	Needs scaffolding

Only the bottom row scales past one person. The first three are why your team is stuck.

Tier 1 of 4

Ephemeral , the napkin.

01·copy · paste · prompt

Context that lives only inside one chat. You paste it in, the model uses it, the chat ends, the memory dies.

You're using this if you use…

ChatGPT memory
Claude Projects
Custom GPTs
Cursor Rules
system prompts
"here's my context, here's my question…"

Why it's useful

Zero setup. Works in any tool.
Immediate. Fine for one-off tasks.
You always know exactly what the AI saw.

Why it breaks

Manual every single time.
Doesn't survive a new chat.
Zero team sharing. Zero version history.

Use it for A one-off prompt where the context fits in a paragraph and the answer fits in a tweet. Not your business.

Tier 2 of 4

Siloed Static , the genius notebook.

02·local · markdown · single device

A real, persistent knowledge vault, but it lives on one machine. You can use it. Nobody else can.

You're using this if you use…

Obsidian
LogSeq
Apple Notes
Notion (personal)
a Desktop folder of .md files
Claude Code + local repo

Why it's useful

Real persistence. Survives new chats.
Markdown is human + AI readable.
Fast, local, private.

Why it breaks

Trapped on one laptop. Other laptop blind.
Your VA, supplier, designer can't see it.
Cloud agents (n8n, background Claudes) can't read it.
Sync via Dropbox = merge-conflict hell.

Use it for Solo founder, single workstation, content you genuinely never want to share. The moment you have a teammate, this is broken.

Tier 3 of 4

Vector Space , the black box.

03·embeddings · similarity · RAG

Your text gets converted into numbers and stored in a database that retrieves by meaning. Powerful. Engineer-shaped.

You're using this if you use…

Pinecone
Chroma
Weaviate
OpenAI Assistants (file search)
LlamaIndex
raw RAG pipelines

Why it's useful

Scales to millions of documents.
Semantic search, not keyword.
AI-native. Built for retrieval.

Why it breaks

Humans cannot read or edit it directly.
Debugging "why didn't it retrieve X?" is brutal.
Re-embed on every schema change.
Engineer required. Bill scales with usage.

Use it for You have 10,000+ documents, a dedicated engineering team, and a search problem. Not a memory problem.

Tier 4 of 4 · the answer

Dynamic Shared , the team brain.

04·cloud markdown · MCP · team-visible

A cloud-hosted markdown vault exposed to every AI via MCP. Humans read it. AIs write it. Your whole team sees the same brain.

You're using this if you use…

InventoryHero MCP
GitHub repo + Claude Code
Letta
OpenMemory
a custom MCP on top of a shared Drive

Why it wins

Humans and AIs read and write.
Whole team sees the same brain.
Version controlled. Diffable. Auditable.
Works from mobile, voice, CLI, any agent.
Curates context, plays nicely with the U-curve.

The cost

Initial scaffolding (the Handbook).
New paradigm, new habits.
Either an afternoon (Road 1) or 5 minutes (Road 2).

Use it for Any team of 2+ humans. Any business running 2+ AI agents. Which is, basically, all of you.

The Local Silo Problem

Brilliance trapped on one drive is brilliance wasted.

"If your context lives in a brilliant Obsidian vault on one laptop…

it's invisible to your ops manager,
your TikTok Shop VA,
your supplier in Shenzhen,
and the 10 Claudes running your product research."

A real story · last week

A CEO of a $40M revenue brand told me…

"I just got it set up.

I have it on my laptop at home.

I have it on my laptop at work.

And they don't talk to each other.

And it's so frustrating."

, anonymous 8-figure Amazon seller

Then I asked: do you share it with your team? He said: no.

Different ≠ Better

Same rule as your products.

You don't need a better memory system.
You need a different one.

One that lives in the cloud.
One that humans can read.
One that AIs can write.
One every team member can see.

Local vaults are the
PDFs of AI memory.

It's time to kill them.

The 18-Month Org Chart

The perfect e-com team in 18 months.

humans

AIs

shared brain

Every company will be an AI company. Every employee will work alongside AI.

Jensen Huang founder & CEO, NVIDIA

(paraphrased from GTC 2024 keynote)

Module 4 · 9 minutes · The Framework

The
AI Employee
Handbook.

The named framework for shared team memory. This is the part you'll leave with.

The Reframe

Knowledge base.
Wiki.
Second brain.

Your AIs aren't readers.
They're workers.

What they need is an
Employee Handbook.

What goes in the Handbook

Five sections. 10 to 100 markdown files.

01 · Identity

Who you are

Brand, mission, voice, customer avatar.

02 · Catalog

What you sell

ASINs, USPs, keyword clusters, pricing logic.

03 · Marketing

How you market

Listing rules, image stacks, email voice, TikTok angle, Shopify brand.

04 · Ops

How you operate

Org chart, suppliers, freight, lead times, PPC rules.

05 · Team

Who's on the team

Humans and active AI agents.

Not 10,000 pages. 100, because nobody reads 10,000.

The Actual Folder Structure

Drop this in a folder. Done.

core-company-wiki/
├── handbook_overview.md
├── brand/
│   ├── brand_guidelines.md       # voice · banned phrases · hero copy
│   └── customer_avatars.md       # e.g., "organic-leaning millennial moms"
├── products/
│   ├── catalog_index.md          # SKU → ASIN map
│   ├── [ASIN]_listing.md         # title · bullets · A+ · keyword clusters
│   └── [ASIN]_image_stack.md     # main image rules · lifestyle · infographics
├── marketing/
│   ├── ppc_rules.md              # ACoS targets · dayparting · defensive bids
│   ├── email_playbook.md         # Klaviyo flows · voice · send cadence
│   ├── tiktok_shop_playbook.md   # creator briefs · hook formulas · hashtags
│   └── shopify_brand.md          # DTC tone · hero pages · CRO rules
└── operations/
    ├── org_structure.md          # humans + active AI agents
    ├── supplier_directory.md     # factories · MOQs · lead times
    └── reorder_rules.md          # when InventoryHero should trigger POs

80/20 Rule for Memory

Scaffolded, not rigid.

80%

Scaffolded

Folder structure + file names are fixed. AIs always know where to look.

20%

Fluid

Inside each file, the AI writes freely. Markdown adapts as your business does.

Rigid databases die.
Markdown adapts.

Progressive Elaboration

Every file knows how complete it is.

# Customer Avatars

[ Sparse, pending Socratic fill ]

▶The AI will literally come ask you what's missing.
▶You answer 4 questions.
▶The file fills itself out.
▶You never sit down to write a wiki. You answer questions over time.

This isn't a wiki.

It's onboarding ,
for every AI you'll ever hire.

In the next 12 months, you'll have AI coworkers in Slack.

Dario Amodei CEO, Anthropic

That day is today. The question is, do your AI coworkers know your brand?

Module 5 · 7 minutes · Power-User Route

Karpathy + GitHub.
Free. Advanced.

A memory layer with zero monthly cost. For the comfortable-in-terminal crowd.

The Fork in the Road

Two paths. Both work. Pick yours.

1 The Power-User Path

Free
Karpathy-style LLM repo
GitHub as your memory store
Total control · total ownership
Requires terminal comfort

We're doing this one first.

2 The Easy Button

5-minute hosted MCP setup
Non-technical team friendly
Mobile + voice friendly
Same Employee Handbook framework
$Hosted (free in beta)

Module 5, coming up next.

If "git push" sounds scary, relax, watch this section, then act on Road 2.

The Karpathy Pattern

From `nanoGPT` & "Software 3.0", applied to your brand.

repo/
├── README.md                  # the map
├── prompts/                   # reusable instruction templates
├── context/                    #  your Employee Handbook lives here
│   ├── brand/
│   ├── products/
│   ├── marketing/
│   └── operations/
├── data/                       # raw inputs · SP-API exports · reviews · sales
└── outputs/                    # what the AI generated · drafts · briefs · ads

Plain text. Git-tracked. Diffable. Forkable. Free forever.

Why GitHub Works

GitHub is shockingly good at this.

Free for private repos.
Built-in version history, never lose context.
Pull requests = paper trail for every memory change.
Collaborators = instant team access.
Works with Claude Code · Cursor · Codex · Gemini CLI.
git push from one laptop → git pull from another → synced.

LIVE DEMO

GitHub as memory, in 7 steps.

Create a private repo: my-brand-handbook
Drop in the Karpathy folder structure.
Open it in Claude Code (or Cursor).
"Read my brand_guidelines.md and write 5 TikTok Shop hooks for ASIN B0XXXXXXX."
Claude writes the hooks → I review → git commit → git push.
Switch to my second laptop → git pull → the memory is there.
Add my VA as a Collaborator → she has the same brain.

Is Road 1 right for you?

Be honest with yourself.

This IS you if…

You're comfortable in a terminal.
You use Cursor / Claude Code daily.
You want $0/mo memory costs.
You want a full audit history.
You sell products AND ship code.

This is NOT you if…

You've never run git push.
Your team is mostly VAs & creatives.
You want this working in 5 minutes.
You want to use it from your phone.

The Honest Catch

Truth time.

You'll spend an afternoon setting it up.
You'll Google "merge conflict" at least once.
Your VAs probably won't touch it.

That's fine, most $1M to $10M Amazon teams want…

Road 2. (next section →)

Software is no longer just written, it's grown.
Memory and context are the soil.

Andrej Karpathy adapted from "Software 3.0"

(paraphrased)

Module 6 · 8 minutes · The Easy Button

Hosted MCP.
5 minutes flat.

Same framework. Zero setup. Mobile + voice friendly. Built for the rest of your team.

The Architecture

Two tools. One vault.

Claudedesktop · code

Geminimobile · CLI

Cursoreditor

ChatGPTdesktop

reads & writes via MCP

InventoryHero.ai MCPread_memory · write_memory

Internal AI Managerstructures · routes · de-duplicates

Secure Markdown Cloud Vaultyour URL · 4-digit PIN · shareable

The Whole API

Two tools. That's it.

Tool 01

read_memory()

"Get me what you know about X."

Tool 02

write_memory()

"Here's something new, file it."

Same Employee Handbook framework.
We do the structuring behind the scenes.

Auto-Population

You don't fill 100 blank files.

If you signed up at slide 17, this is already running.

Source · Live

Your Amazon SP-API

→ populates catalog_index + per-ASIN listing files.

Source · Live

Keepa product history

→ pricing context, BSR trends, competitor moves.

Source · You

Uploaded brand docs

→ brand_guidelines.md + customer_avatars.md.

Source · AI

4 Socratic questions

→ fills in the gaps the data can't see.

LIVE DEMO · #1

The mobile vignette. I'm on a run.

Headphones in. Phone in pocket. No laptop in sight.

Voice memo to Gemini. It pulls my brand voice. Pulls my product context. Writes three TikTok hooks. While I run.

The unlock: the memory layer is independent of which AI app you use.

9:42 AM · Trail

You · voice Hey Gemini, draft 3 TikTok Shop hook ideas for my B0XXXXXXX product. Use my brand voice.

For context, read memory.delmarave.io, PIN 1234.

Gemini · 0.8s Pulled brand voice (warm, witty, organic-leaning). Three hooks coming up, drop 1 leads with a problem, drop 2 with a stat, drop 3 with a POV…

LIVE DEMO · #2

The write path. Claude logs it for the team.

Claude · agent log

▸ read_memory("catalog/B0XYZ")
↳ found: hero ASIN, margin playbook, current price.

▸ fetch_sp_api("inventory/B0XYZ")
↳ supplier cost: +8.1% vs prior order.

▸ write_memory("catalog/B0XYZ", patch)
+ flagged: sourcing_cost_change
+ recommended: test +$0.50 price lift
+ team_visible: true

"I'm updating your catalog.
Your hero ASIN's sourcing cost jumped 8%.
Your margin playbook says test a price lift before absorbing it.
I've logged this for the team."

Watch the markdown file update, live on screen.

Live audience reveal

Half the room is already running.

If you signed up at slide 17, you're already running.
→Half signing up now will be live by slide 60.
→The other half: by end of Q&A.
8-figure sellers, you'll probably break us a little. We'll wear it like a badge.

Last chance to sync

inventoryhero.ai/signup

Every company will deploy thousands of agents alongside their employees.

Marc Benioff CEO, Salesforce

(paraphrased from Dreamforce 2024)

What we covered

The whole talk, in four pieces.

AI memory ≈ human memory.
Working vs long-term.
Local vaults are dead.
They're the PDFs of AI memory.
The AI Employee Handbook. ○ 10 to 100 markdown files ○ Scaffolded, not rigid ○ Progressive elaboration
Two deployment paths. ○ Karpathy + GitHub, free, advanced ○ Hosted MCP, easy, hosted

Outcome: every AI in your business reads from the same brain.

The Amazon sellers who win the next 5 years
won't have the smartest AIs.

They'll have the
best-onboarded ones.

It's not about man versus machine. It's about augmenting human capabilities through AI.

Satya Nadella chair & CEO, Microsoft

Build it with me · 100% free

AI Hackathon.

Tuesday · June 2 · 9 AM PT / 12 PM ET / 5 PM London

We'll walk through:

Spinning up your Employee Handbook from scratch.
Wiring up the GitHub power-user path (Road 1).
Installing the InventoryHero MCP (Road 2).
Live Q&A with my dev team.

100% free. Bring your laptop and a cup of coffee.

Save your seat

inventoryhero.ai/hackathon

The other thing I'm building

InventoryHero.ai

The agentic inventory operating system for Amazon.

Talks to your SP-API in plain English.
Detects stockouts, overstock, reorder risk.
Acts on it, not just reports it.
Built on the memory layer you just learned about.
Multi-channel ready: Amazon → Shopify → TikTok Shop.

Beta open. Free to sign up. Limited slots.

Join the beta

inventoryhero.ai

Q&A

That's the whole framework. What did I miss?

Andrew@TitanMembers.com inventoryhero.ai/hackathon Tuesday · 9 AM PT

Build YourLLM Team Wiki

Andrew Erickson

What I'm building now

Aliza & Arlo

Let's talk business.

Blow your mind.

How well you understand the next 60 minutes will determine whether your business is running 10 AIs in 12 months… or being run over by them.

The next 60 minutes, in four pieces.

Your AIs have amnesia.

AI memory is broken, especially for Amazon teams.

You can't out-prompt this.You have to out-structure it.

You're conducting a symphony where everyone has different sheet music.

Sign up now, your account syncs while I talk.

Three reasons to sign up while you listen.

The Anatomy ofAI Memory.

Don't think "vector embeddings."Think about how YOU remember.

The Context Window = Working Memory

Vector Embeddings + Retrieval

How ContextActually Works.

Every chat has a token budget.

AI recall is a U-curve, not a flat line.

Your data is the thin slice. Most of the budget is overhead.

Never dump. Always curate.

The Hierarchy ofMemory Tools.

Four tiers. Three of them fail your team.

Ephemeral , the napkin.

Siloed Static , the genius notebook.

Vector Space , the black box.

Dynamic Shared , the team brain.

Brilliance trapped on one drive is brilliance wasted.

A CEO of a $40M revenue brand told me…

Same rule as your products.

The perfect e-com team in 18 months.

TheAI EmployeeHandbook.

Knowledge base. Wiki. Second brain.

Five sections. 10 to 100 markdown files.

Drop this in a folder. Done.

Scaffolded, not rigid.

Every file knows how complete it is.

Karpathy + GitHub.Free. Advanced.

Two paths. Both work. Pick yours.

From nanoGPT & "Software 3.0", applied to your brand.

GitHub is shockingly good at this.

GitHub as memory, in 7 steps.

Be honest with yourself.

Truth time.

Hosted MCP.5 minutes flat.

Two tools. One vault.

Two tools. That's it.

You don't fill 100 blank files.

The mobile vignette. I'm on a run.

The write path. Claude logs it for the team.

Half the room is already running.

The whole talk, in four pieces.

AI Hackathon.

InventoryHero.ai

Build Your
LLM Team Wiki

How well you understand the next 60 minutes will determine whether your business is running 10 AIs in 12 months…
or being run over by them.

You can't out-prompt this.
You have to out-structure it.

The Anatomy of
AI Memory.

Don't think "vector embeddings."
Think about how YOU remember.

How Context
Actually Works.

The Hierarchy of
Memory Tools.

The
AI Employee
Handbook.

Knowledge base.
Wiki.
Second brain.

Karpathy + GitHub.
Free. Advanced.

From `nanoGPT` & "Software 3.0", applied to your brand.

Hosted MCP.
5 minutes flat.