Edition 340 | April 8, 2026 The Dyslexic AI Newsletter by LM Lab AI

Reading Time: 9 minutes Listening Time: 13 minutes

I finished writing Edition 339 and I did not close my laptop.

I sat there staring at Karpathy's system and Chappy Asel's five-layer framework and a thought hit me that I could not shake.

I already have most of this.

I am writing this from the waiting room at EmpowerED Educational Psychology Services in Sebastopol. My daughter Makena is on the other side of the wall getting assessed for learning differences. While she does her thing, I am doing mine: vibe coding a custom piece of software for her homeschool curriculum on one screen and writing this newsletter on the other.

If you have been here since Edition 325 ("My 14-Year-Old Daughter Just Proved Me Wrong"), you know Makena is part of this story. She is the reason I build half of what I build. And right now, in this waiting room, I am watching the same pattern play out that I want to talk to you about today.

I am taking pieces that already exist. Assessment data. Curriculum frameworks. AI tools I have been building for months. And I am looking for the thing that connects them into something that improves over time.

The loop.

My Single Source of Truth. My cognitive profile. My Claude Skills. The frameworks I have been building for three years.

Four layers. Knowledge, memory, context, skills. All built. All working.

But not the fifth. Not the loop. Not the part where the system gets better every time it runs. The part where outputs feed back into inputs and everything compounds.

I have been building the pieces. Now it is time to connect them.

And I want to do it in the open. With you. So you can build yours alongside me.

If you are new here, let me get you caught up fast.

In Edition 329 ("Building Your Second Brain"), I introduced the Single Source of Truth. One document that tells your AI who you are, how you think, what you are working on, and what kind of output you need. You paste it at the top of every conversation so your AI starts as a partner, not a stranger.

In Edition 332 ("A Year Ago, I Was in a Hospital Bed"), I introduced the Cognitive Balance Model. Three phases: Human Initiation (you set direction), AI Expansion (the AI does deep work), Human Integration (you make the final call). Each phase scores 1 to 5 on the Human Guidance Index (HGI) for a total of 3 to 15.

In Edition 335 ("Your Brain Has a Profile Now"), we went deeper. You built a cognitive profile. A map of how your brain works with AI. Not just preferences. Patterns.

In Edition 339 ("Your AI Just Forgot Everything. Again."), we looked at the research. Karpathy's knowledge base system. Chappy Asel's five-layer stack. And we saw that what we had been building lined up with what the top engineers in AI were discovering independently.

Today we design the loop. And then we run it together.

Let me demystify this before we build it.

A self-improving loop is not some sci-fi concept. It is three steps repeated on a schedule.

Step 1: Work with your AI. Do your normal thing. Write. Brainstorm. Solve problems. Whatever your Tuesday looks like.

Step 2: Review what happened. At the end of the week, look at what worked and what did not. Where did the AI nail it? Where did it miss? What did you have to correct three times?

Step 3: Update your source document. Take what you learned and feed it back into your Single Source of Truth or cognitive profile. The next session starts smarter than the last one.

That is the whole loop. Work, review, update. Repeat.

The magic is not in any single step. It is in the repetition. Every cycle makes the next one better. This is what Karpathy's system does with code. We are going to do it with cognition.

Claude now has memory for all users, including the free tier. That is a real shift from even six months ago. But here is what most people miss: Claude's built-in memory is a summary, not a system. It captures preferences and facts. It does not capture your thinking patterns, your decision-making process, or the context behind why you work the way you do.

That is what your Single Source of Truth is for. And that is what we are going to make self-improving.

Step 1: The Weekly Debrief Prompt

At the end of each week, open a new Claude conversation. Paste your Single Source of Truth at the top. Then paste this:

That last line is important. Human Integration. You make the call. The AI proposes. You decide. That is a Cognitive Balance Model score of about 13 out of 15.

Step 2: The Memory Sync

After you approve the updates to your Single Source of Truth, paste this follow-up:

Save those three anchors somewhere you can find them. Top of your Single Source of Truth. A sticky note on your monitor. Wherever works. They become your bridge between sessions.

If you are a paid Claude user, you can also go into Settings and manually edit your memory to reflect these anchors. Claude's memory system updates every 24 hours automatically, but sometimes the automatic summary misses the nuance. Manual edits give you control.

Step 3: The Structured Forgetting Pass

Once a month, run this:

This is the structured forgetting concept from Edition 339. Old information does not just sit there. It actively confuses the AI. A clean document performs better than a comprehensive one.

ChatGPT handles memory differently than Claude. It stores discrete memory entries that persist across conversations automatically. You can view them in Settings under Personalization. It also has Custom Instructions, which function like a simplified version of the Single Source of Truth.

The advantage: ChatGPT's memory is always on. You do not have to paste anything. It just remembers.

The disadvantage: you have less control over what it remembers and how it organizes that information. It can get noisy fast.

Here is how to build the same loop.

Step 1: The Weekly Audit Prompt

ChatGPT will show you its stored memories and propose edits. You review and approve. Same Human Integration gate.

Step 2: Custom Instructions Refresh

ChatGPT's Custom Instructions are your Single Source of Truth equivalent. They are limited in length, so precision matters. After your weekly audit, update your Custom Instructions with this prompt:

That character limit forces clarity. Which, honestly, is a feature, not a bug. Constraints make you choose what actually matters.

Step 3: The Cross-Platform Bridge

Here is something most people are not doing yet. If you use both Claude and ChatGPT (and I do), your self-improving loop should feed both systems.

After your weekly review in either tool, export the key updates and paste them into the other. Claude added a memory import tool in March 2026 that makes this easier. You can export your ChatGPT memories using a prompt and import them directly into Claude's memory settings.

The goal: both tools should have the same map of who you are. Not identical documents. The same understanding.

I am being straight with you. I have not run this loop yet. I designed it this week based on the research from Edition 339, the frameworks I have been building for three years, and the way memory works in both tools right now.

But I know enough to have some predictions. And I want to put them on the record so we can check them together.

I think the AI will catch patterns I miss. I already know from building my cognitive profile in Edition 335 that the AI sees things in my workflow that I am too close to notice. A structured weekly review should surface more of those.

I think structured forgetting will be harder than it sounds. I have framework descriptions in my source document from eighteen months ago that I have not touched. Some are probably outdated. Some have probably evolved into something better. But deleting work feels like losing work. I expect the archive option will get a lot of use early on.

I think Claude and ChatGPT will see different things. Claude's strength is structural analysis. ChatGPT's strength is persistent detail recall. Running the loop in both should give a more complete picture than either one alone.

I think the compound effect will be slow at first. The research says these systems improve exponentially over time. But the first few cycles are mostly about getting the foundation clean. The gains come later.

I will report back. That is a promise. This is an experiment we are running together, and I will share what works, what does not, and what surprises me.

I keep coming back to this.

Every framework in the self-improving stack maps to something neurodivergent thinkers already do. We talked about that in Edition 339. But the loop itself? The review-and-update cycle? That is something dyslexic minds are uniquely good at.

We iterate. Constantly. We do not get things right the first time. We never have. So we developed the ability to look at what we produced, figure out what is off, and try again with better information.

That is literally what the self-improving loop does.

The engineers who built these systems had to design the ability to iterate. We were born with it.

In Edition 330 ("What Do Problem Solvers Do When AI Solves All the Problems?"), I asked what happens when the hard problems get automated. The answer is becoming clearer every week. The hard problem is not solving tasks. It is building systems that get better over time. And the people who are best at that? The ones who have been iterating their way through a world not designed for them since kindergarten.

In Edition 334 ("The Data Is In"), we saw the Anthropic labor market study that showed cognitive flexibility as job security. The self-improving loop is cognitive flexibility in action. You are not locked into one way of working. You review. You adapt. You update. That is the skill.

Three things. This week.

1. Run Your First Weekly Debrief

Pick either Claude or ChatGPT. Paste your Single Source of Truth (or Custom Instructions). Run the weekly debrief prompt from above. Do not overthink it. Just see what comes back.

If you do not have a Single Source of Truth yet, start with the memory map prompt from Edition 339. Build the document first. Then start the loop next week.

2. Set a Calendar Reminder

The loop only works if you actually do it. Set a recurring 15-minute block. Friday afternoon. Sunday morning. Whatever works for you. Call it "AI Review" or "Cognitive Tune-Up" or whatever will make you actually show up.

The system that gets better every cycle only works if there is a next cycle.

3. Try the Cross-Platform Sync

If you use both Claude and ChatGPT, spend 10 minutes syncing them. Export your ChatGPT memories. Import them into Claude. Or just copy the key points from one tool's review into the other tool's source document. The goal is not perfection. It is making sure both partners have the same map.

Here is the bottom line.

Karpathy ran 700 autonomous improvements in two days on a code training system. We are not doing that. We are doing something smaller and, I would argue, more important.

We are building a system where your AI gets better at understanding you every single week. Not through some automated pipeline. Through a deliberate, human-guided, fifteen-minute review cycle.

That is the Cognitive Balance Model at its best. You initiate. The AI expands. You integrate. And next week, the whole thing starts from a higher baseline.

If you are a Cognitive Partner Member, I am going to be running this experiment in real time and sharing everything I learn with you first. The wins, the failures, the adjustments. I will also be building a pre-made template for the full loop: Single Source of Truth template, weekly debrief prompts, monthly forgetting pass, cross-platform sync guide. Founding members get early access.

If you are not a member yet, you have everything you need in this edition and Edition 339 to start today. The prompts are right here. The framework is right here. The only thing missing is fifteen minutes on a Friday.

50 founding spots. $19 a month. Locked forever. Your AI should get smarter every week. Not start over every session.

Edition 341: Topic coming soon. Stay tuned.

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

Dictated, not typed. Obviously.

🔁 The self-improving loop is three steps: work with your AI, review what happened, update your source document. Repeat weekly.

🛠️ Full prompts included for both Claude and ChatGPT: weekly debriefs, memory syncs, structured forgetting passes, and cross-platform bridging.

🤖 Claude and ChatGPT handle memory differently. Claude gives you more structural control. ChatGPT remembers details automatically. Using both together gives the fullest picture.

🧠 I have not run this loop yet. I designed it this week and I am starting it alongside you. I will report back with real results. No made-up stories.

♻️ Dyslexic thinkers are already wired for iteration. The self-improving loop is what we have been doing our whole lives, formalized.

🔧 Three things to do this week: run your first debrief, set a recurring calendar block, and sync your tools if you use more than one.