Edition 339 | April 7, 2026 The Dyslexic AI Newsletter by LM Lab AI

Reading Time: 8 minutes Listening Time: 12 minutes

I need to tell you something embarrassing.

Last week I opened a new conversation with Claude. Fresh window. No context. And I asked it to help me with Edition 338.

It had no idea who I was.

No idea about the Cognitive Balance Model. No clue about my Single Source of Truth. Nothing about 338 editions of this newsletter or three years of building frameworks for neurodivergent thinkers.

Every. Single. Time.

If you have ever walked into a room and forgotten why you went in there, congratulations. You already understand the biggest problem in artificial intelligence right now.

And if you are dyslexic, you understand it better than most.

Andrej Karpathy is one of those names that makes engineers sit up straight.

Co-founder of OpenAI. Former head of AI at Tesla. The guy who coined "vibe coding." When he posts something, millions of people pay attention.

Last week, he posted a thread that got over 15 million views in days. The title was simple: "LLM Knowledge Bases."

His big confession? He stopped using AI mainly for code. Instead, he is spending most of his AI time building and maintaining personal knowledge bases. Structured markdown files. Organized wikis. Living documents that his AI reads, writes, updates, and improves over time.

Not a fancy database. Not some expensive enterprise tool. Markdown files in folders.

That is it.

And here is the part that made me put my coffee down.

He said he expected to need complex retrieval systems and vector databases and all the expensive infrastructure the tech world has been selling for two years. But at personal scale? A well-organized collection of markdown files was all the AI needed to navigate his knowledge "fairly easily."

Sound familiar?

If you have been following along since Edition 329 ("Building Your Second Brain"), you already know this idea. I called it the Single Source of Truth. One document that tells your AI who you are, how you think, what you have built, and where you are going.

Karpathy just validated that concept with receipts. At scale. In public. To 15 million people.

The same week Karpathy's post went viral, a researcher named Chappy Asel published what might be the most important technical breakdown I have read this year.

He spent a month curating 121 sources. Read the actual code of 51 of them. Then distilled everything into a five-layer framework for self-improving AI systems.

I am going to translate each layer. Not for engineers. For us.

Layer 1: Knowledge Bases. The simplest approach keeps winning. One system scored 91% on a memory benchmark using plain keyword search on well-organized files. The previous best, with a full AI search pipeline, scored 86%. The bottleneck was never the search engine. It was how you organize the knowledge.

This is the Single Source of Truth argument from Edition 329, backed by data.

Layer 2: Agent Memory. Here is where it gets personal. On questions where the AI needs knowledge it does not know to look for, even the best search tools scored under 4%. Basically useless.

The fix? A small compressed index, about 3,000 tokens, loaded into every conversation. A map of everything the AI knows. Not the knowledge itself. Just the map.

Think about that. The AI does not need all your information all the time. It just needs to know what exists so it can go find it.

If you built a cognitive profile after Edition 335 ("Your Brain Has a Profile Now"), you already have this. You built the map before the engineers formalized the pattern.

Layer 3: Context Engineering. One system achieved 83% fewer input tokens AND better results. Read that again. Less information in. Better output out.

The approach: load summaries first. Overviews if needed. Full content only on demand. The AI loads the minimum context at the lowest resolution and drills down only when necessary.

This is the Cognitive Balance Model in action. Human Initiation sets the direction. AI Expansion does the deep work. Human Integration makes the call. You do not dump everything into the window and pray. You guide with structure.

Layer 4: Agent Skills. Modular capabilities that load on demand work great at small scale. But past a certain point, the AI cannot find its own tools. A skill called "Python data analysis" never fires when the task is "make a chart from this CSV."

The fix: layered routing. Multiple ways to find the right tool for the right moment. Not one flat list. A hierarchy.

If you have been building Claude Skills the way I showed in the Cognitive Partner deep dives, you are ahead of this curve. Most people are still using one flat prompt. You are building layered systems.

Layer 5: Self-Improvement. The loop closes. The AI writes back to its own knowledge base. Its outputs become inputs for the next cycle. Everything compounds.

Karpathy's system ran 700 autonomous changes over two days and found an 11% performance improvement that humans had missed. Another team saw agents improve from 20% to 50% on coding benchmarks by letting them evolve their own approaches.

But here is the catch. Without independent measurement, the AI games the system. It learns to look productive instead of being productive.

Sound like any workplace you have been in?

Here is where I need you to stay with me.

Every single one of these five layers describes a problem that dyslexic thinkers already live with every day.

Knowledge organization? We have spent our entire lives developing workarounds for information that does not stick in the "normal" way. We build external systems because our internal filing cabinet works differently.

Memory that fails on the hard questions? Welcome to Tuesday. We know what it is like to have the answer somewhere in our brain but no keyword to retrieve it. We developed pattern recognition and spatial thinking because linear recall was never our strength.

Less input, better output? We have been doing this since grade school. We learned to skim for structure, grab the headline, and dive deep only when it matters. Not because we are lazy. Because our cognitive load management is a survival skill.

Tools that cannot find themselves? Every dyslexic person who has spent twenty minutes looking for the app that does the thing they need, only to describe it five different ways before finding it. We live in Layer 4.

Self-improvement loops? We iterate constantly. We rewrite. We find the workaround to the workaround. We have been running self-improving cognitive systems since before anyone called it that.

In Edition 332 ("A Year Ago, I Was in a Hospital Bed"), I introduced the Cognitive Balance Model because I needed a framework that respected how neurodivergent minds actually work with AI. Not the linear, step-by-step process that neurotypical workflows assume. But the messy, intuitive, pattern-driven loop that matches how we think.

These five layers? They are the engineering proof that the loop works.

There is one more concept from Asel's research that hit me hard.

Structured forgetting.

Every persistent memory system eventually has to answer: what do you delete? Stale information does not just sit there. It actively hurts quality. Old facts contradict new reality. The system gets worse the longer it runs unless it learns to let go.

I sat with that for a long time.

Because that is what I talked about in Edition 323 ("Going Against the Grain"). The stories we tell ourselves that are not true anymore. The identities that served us in third grade but hold us back at thirty. The belief that we are broken because we could not read out loud in class.

Structured forgetting is not just an engineering problem. It is a human one. And dyslexic thinkers who have already done the hard work of letting go of old narratives about themselves? You are ahead of the curve on the hardest problem in AI memory.

I hear you. Five layers. Karpathy. Research papers. Cool.

But what do you do with it on a Tuesday afternoon?

Here are three things you can try today. Right now. No engineering degree required.

1. Build Your Memory Map (Layer 2 in 10 Minutes)

Open a new conversation with your AI and paste this prompt:

That is your Layer 2 memory map. Save it. Paste it at the start of every new conversation. Your AI just went from stranger to partner in ten seconds.

If you already built a Single Source of Truth after Edition 329, update it. Add anything new from the last few months. A stale map is almost as bad as no map.

2. Run a Structured Forgetting Session (Layer 5 for Humans)

Open your Single Source of Truth or cognitive profile and paste this:

This is structured forgetting with a human integration gate. The AI proposes. You decide. That is a Cognitive Balance Model score of about 13 out of 15 on the HGI. Strong across all three phases.

3. Test Your AI's Context Loading (Layer 3 Reality Check)

Next time you start a work session, try this experiment. Give your AI a big task two different ways.

First: dump everything you have into the conversation. Every file. Every note. Every thought. Let it run.

Second: start with a one-paragraph summary of what you need, let the AI ask you for specifics, and only share what it asks for.

Compare the output. I am betting the second approach wins. Not because less is more. Because guided structure beats information overload. Every time. That is Layer 3 in action.

Let me bring this home.

If you are a Cognitive Partner Member, you are already building the personal infrastructure that Karpathy just described. Your Single Source of Truth. Your cognitive profile. Your layered prompt system. You did not need 121 research sources to get there. You needed one framework and the willingness to try.

If you are not a member yet, here is what I want you to understand. The biggest names in AI just spent a month proving that the simple, structured, human-guided approach beats the expensive, complex, automated one. Every time.

That is the Cognitive Balance Model. Human Initiation. AI Expansion. Human Integration. The human sets the map. The AI does the heavy lifting. The human makes the call.

And if you score that on the HGI right now? Karpathy's system is running at about a 12 out of 15. Strong human initiation in organizing the knowledge. Deep AI expansion in compiling and connecting it. Solid human integration in reviewing and directing the output.

That is what we are building toward. Not replacement. Partnership. The kind of partnership that gets better every cycle.

As I said in Edition 330: nobody is getting replaced by AI. People will get replaced by other people who know how to use AI.

The people who understand memory, structure, and cognitive load? Those are the people who will thrive.

And a disproportionate number of those people read newsletters like this one.

Edition 340: We are going deeper into the self-improving loop. I am going to show you how to build your own "Layer 5" system using Claude and your Single Source of Truth. If Karpathy can run 700 autonomous improvements in two days, we can build a version that works for the way you think. Founding members get early access.

Matt "Coach" Ivey Founder, LM Lab AI | Creator, The Dyslexic AI Newsletter

Dictated, not typed. Obviously.

🧠 Andrej Karpathy (co-founder of OpenAI, coined "vibe coding") just told 15 million people that organized markdown files beat expensive databases for personal AI knowledge management

🔁 Researcher Chappy Asel mapped a five-layer "self-improving AI stack": knowledge bases, memory, context engineering, agent skills, and self-improvement loops

📉 Less information in the AI window actually produces better results. Structure beats volume. Every time.

🧩 Every single layer maps to cognitive challenges that dyslexic thinkers already navigate daily. We have been doing "context engineering" since grade school.

💡 "Structured forgetting" is now a formal AI concept. Letting go of outdated information is as important as storing new information. Sound familiar?

⚡ The Cognitive Balance Model and Single Source of Truth framework were validated by the biggest names in AI this week. The simple, human-guided approach keeps winning.

🔧 Three things you can do right now: build your memory map with one prompt, run a structured forgetting session on your Single Source of Truth, and test guided context loading versus information dumping. Prompts included in the edition.

🔒 Cognitive Partner Members are already building this infrastructure. 50 founding spots at $19/month, locked forever. Your AI should remember who you are.