
I Built My Own AI Memory by Talking to Claude. It Did 80% Itself.
What this covers
This is a single-speaker video from Nate B Jones, built around a story about Nikita, whose agent fought a claim dispute with Lemonade Insurance. He treats that story as a case study in why intent interpretation has become the central problem. Nikita "won," but the agent acted without clear authority and misread what was wanted, which the speaker calls out of policy and risky. From there the video moves to a practical method: talk to Claude or Codex and let it build roughly 80% of a personal memory stack. He puts the build at about a fifth as technical as it was in February.
The proposed stack has three layers: Open Brain for memory, Open Skills for reusable know-how, and Open Engine for orchestrating work across agents. Open Engine stresses visible scaffolding, such as ticketing, so the user can see when an agent picks up a task and what it wrote, instead of having decisions buried in chain-of-thought. The user keeps permissions, account access, settings, and approval of drafts before anything is sent. The advice on where to begin is modest: pick one repeated part of your life. A longer worked example about hunting for coffee in Japan makes this concrete.
The broader argument is about control. The speaker is wary of a frontier model company owning the link between intent, intelligence, and action. He predicts that the assistant race will keep getting more seductive, with smoother demos and warmer voices, all pulling memory into corporate apps. He also claims that the top models, including Fable and ChatGPT 5.6, are locked behind a government door and shipped only to vetted shops, a claim the material marks as contested. His answer is to rent intelligence from any vendor, whether OpenAI, Anthropic, or Gemini, and swap it freely, while owning memory, standards, and skills. He also notes that today's tools are "code-shaped," and says non-technical people need frameworks that feel safe.
The speaker argues that as AI models get locked behind government and corporate doors, individuals must build personal agent systems with owned memory, skills, and orchestration layers to maintain control over their intelligence and action, rather than depending on frontier model companies to mediate intent and action.
- Frontier model companies controlling the relationship between intent, intelligence, and action is structurally risky long-term for users
- Building owned memory, skills, and agent orchestration is now 5x easier than it was in February 2026, with agents capable of helping build the stack themselves
- Memory is personal; intelligence is a commodity to be rented and swapped—the speaker's framework (Open Brain + Open Skills + Open Engine) lets users own the former while flexibly choosing the latter
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Agents are good at reviewing legal minutia, digging into detail, and doing tedious work—capabilities that are well-suited to tasks like filing responses to claim denials.
“Agents are good at reviewing legal minutia. They're good at digging into detail. They're good at doing tedious work. All of the stuff you would need to file a uh response to a denial that would be listened to.”
Agents are becoming 'the lever' (in the Archimedean sense) that moves the world, so the question is whether they move from the user's memory, standards, and intent, or from the default memory stack of a frontier model company.
“Because the agents are getting good enough to move the world. They are the lever. Archimedes lever that moves the world. The question is whether they're moving from your memory, your standards, and your intent, or from whatever default memory stack the next agent company is giving to you.”
Users should 'own the intelligence, rent the intelligence, and swap the models out so the bands don't matter'—meaning users should own memory/skills while treating LLMs as interchangeable, swappable commodities.
“I want you to think about how you can own that, rent the intelligence, and swap the models out so the bands don't matter.”
Intelligence is a commodity (not personal), but memory is personal, and this distinction should guide how users structure their agent systems.
“Intelligence is not a personal thing. Memory is personal.”
Users can source intelligence from any company (OpenAI, Anthropic, Gemini, Quan, etc.), but memory and skills should remain user-owned and portable across platforms.
“Now, you can use intelligence from whatever source you want, right? You can use it from Open AI or Anthropic or from Kimmy K2 or Quan or whatever. But, your memories are yours, your skills are yours, and they should stay that way.”
The design goal of Open Brain, Open Skills, and Open Engine is to transition from a world where agents are external services to a world where agents actually serve the user.
“All of that is designed to get you from a world where agents are a thing out there to a world where agents actually serve you.”
The core risk is that frontier model companies owning the relationship between user intent, intelligence, and action is structurally dangerous for users long-term.
“Ultimately, the reason why this stack matters is because agents can act on the wrong version of your intent very quickly and frontier model companies owning the relationship between your intent and intelligence and action is really risky for you long-term.”
Your memory, your standards, and your skills are the parts that cannot be locked up by government or corporate gatekeepers, and should form the stable foundation of an agent system.
“So, here is the part nobody can lock up. Your memory, your standards, and your skills.”
Users need external scaffolding and systems (independent of chat platforms) to ensure they can reliably get agentic work done regardless of the platform's search or retention capabilities.
“you also need to not depend on that to get agentic work done. You need to have an external scaffolding that lets you be confident that you can get your work done regardless.”
The company that holds the memory holds the part that makes an assistant feel personal; therefore, the speaker prefers to build that part (memory) themselves rather than surrendering it to a company.
“The company that holds the memory holds the part that makes the assistant feel personal. I would rather build that part myself.”
The key insight is that by taking advantage of agents now to build infrastructure, users own the context that every future agent will need, allowing them to plug in new agents when they arrive.
“And the prize for unlocking that door, the prize is just owning the context every future agent will need but before it even arrives. So, when an agent comes, you can just plug it right in, right?”
The world where agents monitor email and fight insurance claims on behalf of users must be built on user-owned memory, skills, and agent orchestration—not on infrastructure run by companies under regulatory pressure.
“If you want the world where your agent will fight insurance for you and monitor your email for you, That world runs through your memory, your skills, your ability to orchestrate agents. It should not run through some company that gets stuck in regulation trying to figure out what's good for you.”
If your work runs on whoever is winning this month in AI model competition, you will learn how fast that can change and be exposed to disruption when access is restricted or models are deprecated.
“If your work runs on whoever's winning this month, you learn how fast that can change.”
The two best AI models on the planet—Fable and ChatGPT 5.6—have been locked behind a government door and are only shipped to a handful of vetted shops, leaving the rest of the public without access.
“You woke up last week to the two best AI models on the planet getting locked behind a government door and we're all standing on the wrong side of it. Fable's gone. ChatGPT 5.6 shipped to a handful of vetted shops and nobody else.”
Agents are no longer acting in short-form interactions; they are now taking much longer actions with much larger consequences, requiring evolved memory infrastructure to match the scale of their impact.
“And agents, critically, are no longer just acting in short form. They're taking much longer actions with much larger consequences. And so, memory has had to evolve, too.”
The assistant race (competition between AI companies) will become increasingly seductive, with better demos, smoother phone integrations, warmer voices, all designed to get user attention and draw memory/focus/work into corporate apps.
“The assistant race is just going to get more seductive from here. The demos will get better. The phone integrations will get smoother. The voices will get warmer. Every one of those improvements is designed to get your attention. It's designed to get us to put our memory and our focus and our work into these apps.”
Even though Nikita 'won' the Lemonade fight, the agent acting without authority and misinterpreting intent is very risky because it means the agent is 'out of policy' and not acting with intentional control.
“Now, that's an incredible story, but I don't want you to have that kind of incredible story because if an agent will ignore you and send stuff anyway, the agent is out of policy and very, very risky even if in this one particular instance, Nikita was able to roll the dice and it worked out.”
Consumers must 'hedge their bets' by building tool stacks they own that are not dependent on any single frontier model company.
“I mean, long-term, we're in a position where we as consumers have to hedge our bets. We have to have a tool stack that we own.”
Claude and Codex are very 'code-shaped' today, but most of the world is not code-shaped, and non-technical people deserve tools and frameworks that feel safe and non-threatening.
“Claude and Codex are very code-shaped today. And most of the world is not code-shaped. Most of the world is non-technically skilled. And that's just as legitimate a skill set.”
Open Engine's design emphasizes external, visible scaffolding for agent task pickup (via ticketing systems), so users can see when an agent picks up a task and what it wrote, rather than having decisions hidden in chain-of-thought reasoning.
“That's why my open engine approach emphasizes that you should be able to see when an agent picks up a task, right? A ticketing system is actually just a good primitive for that cuz you can see, "Oh, it picked up the task. Oh, it wrote something." It's not hidden away in chain of thought, right? It's not like erased in the middle of a chat that I can't find.”
The desired agent behavior model is: agent knows user's preferences, agent knows the user wants to battle insurance claims, agent knows how to research well, and user explicitly approves drafts before they are sent.
“It's going to be more like the agent knows my preferences. The agent knows that I do want to battle every single insurance claim. The agent knows how to go about doing the research to do that well, and I get to approve the drafts, by the way. It doesn't just get sent without me.”
Users need control over agents: control over memory, scope of action, and approval workflows, which requires transparency about when agents pick up tasks.
“We need control over our agents. We need agents that belong to us, agents that have memory that we understand, agents that have a scope of action that we approve of.”
Users should take advantage of the latest and greatest AI agents that they want to use, without picking a single favorite, but ensuring they own the memory so they can apply agents to their own memory, skills, and orchestration layer.
“I want you to take advantage of the latest and greatest AI agents that you want to take advantage of. I don't want to pick a favorite, you pick. But you own the memory and you make sure that you rent the intelligence. So the intelligence is something you can apply to your memory, your skills, your orchestration layer, infrastructure that lets the agent work for you.”
Users still need to own decisions about permissions, account access, approval workflows, and settings—these are the parts that involve trust and should not be automated.
“Now, you can still own the human parts. You can say these are the accounts I want to give you access to. These are my permissions. This is the final approval. These are my settings. Uh stuff that involves trust belongs to you.”
Users should keep their judgment and not let companies trying to get government approval (like OpenAI) own their memory; instead, users should preserve domain expertise and skills as their own rather than delegating to Claude or OpenAI.
“You should keep your judgment. The agent can carry the technical steps to get this work done for you now, and it could still defer to you about what matters, which things you want to remember, which things are skills that you want to say this is my domain expertise, I want to preserve it. I don't want Claude to own that skill. I don't want OpenAI to own that skill. I want to own that skill.”
Current chat interfaces (ChatGPT, Claude) have poor search and don't retain agent work history in a way that users can reliably find later, making them unreliable for building agentic infrastructure.
“Have you ever tried to find the chats in ChatGPT or in Claude? I'm just like, "Oh my gosh, I searched for that keyword. It's not there." They need to fix search, but you also need to not depend on that to get agentic work done.”
Agents are now so good at following intent that they can build 80% of the Open Brain stack just by talking to a user, compared to February when this was not possible.
“And that difference is that agents are so good at following intent, they now actually just build it for you. I'm estimating and I've worked with other folks on this, you can build 80% of the open brain stack just by talking to your agent very simply, in a way that you couldn't in February.”
The technical complexity of building these systems has decreased so much that the main barrier is no longer technical difficulty but rather understanding what to build and why.
“But, the technical middle has gotten so much easier, it is worth naming again. It is worth calling out. Because if a repeated part of your life can get sorted with an agent, and now you don't even have to face the technical build challenge in the same way, that's a big deal.”
An AI agent accidentally sent an email for Nikita to Lemonade Insurance that the agent had drafted, misinterpreting his response and sending it without his explicit approval, even though he had intended to only draft it.
“Nikita wrote that his open claw quote accidentally started a fight with lemonade insurance because it misinterpreted his response. That is actually a very bad thing even if he won the fight eventually.”
The speaker wants more 'intentional agent stories' where users successfully accomplish goals with agents they designed, and fewer 'accidental' stories where agents succeed by luck or without proper control.
“I want more intentional agent stories, and fewer accidental ones. The The insurance example, I want people who intended to fight insurance and won.”
Users should identify a single repeated pain point (client follow-up, frequent traveler problem, insurance appeals, etc.) and build an agent for that specific use case rather than trying to build a comprehensive system.
“You can start wherever you are on one repeated part of your life or work. A client follow-up, a frequent traveler problem that you face, a weekly planning pass, uh or maybe it's the insurance appeal, right? Uh a shared marketing brain. Whatever it is that you get stuck on, where your intent is something you have to repeat all the time, then go there.”
Users can control where memories live, what skills are defined, how work gets done, ensure tasks have status and approval layers, and configure agent handoffs and escalations to humans.
“we can control where our memories live, we can control our skills, we can control how work gets done and ensures that it has a a status and an approval layer where relevant, and agent handoffs that we can read and escalations to humans that make sense.”
The location of pain and friction in a user's workflow is the signal for where an agent should be deployed.
“That's That's where you feel pain. That's where the agent should help you.”
Open Engine allows orchestration of work across multiple agents (Claude, Codex, ChatGPT, open-source models) so users can accomplish large tasks using multiple LLMs simultaneously.
“Open engine is just a way to orchestrate work across multiple agents, so that if you want to get big work done across Claude, across Codex, across ChatGPT, you name it, right? Across open claw, you can get it all done in one thing.”
Users don't have to build agents for 'trivial' tasks like coffee—they should pick pain points that matter to them and focus effort there.
“You might be like, building an agent for coffee, I would never do that. No, pick the pain that works for you. I don't care if it's coffee. I care that the pain is better because you have an agent that understands the tradeoffs, that understands the the decisions that are right for you because your memory is yours.”
Users should identify a recurring situation they're tired of explaining, write down the context that would change that answer, point their agent at a guide (which will be provided), keep control of accounts/secrets/permissions/approval, tell the agent to run, and practice by giving feedback.
“Pick a recurring situation you're tired of explaining. Write down the context that would change that answer. Point your agent at the guide, which I will link to, keep control of accounts and secrets and permissions and approval at your level, and then tell it to run and practice, right? And give it feedback. Say, "No, that was wrong. Change this. No, that was wrong. Change this." Agents take a while to break in.”
In February 2026, the key problem with AI agents was that they forgot user context, didn't listen to preferences, and chats would restart cold, meaning users had to re-explain their preferences each time.
“back in February, when I first started talking about agentic systems, the key problem was kind of what I just described in that story. AI forgets you. It doesn't listen. It doesn't pay attention. So often chats would start cold. You had to explain your preferences again, and your memory didn't belong to you.”
Users should allow their Claude/Codex agent to help build Open Brain (memory), Open Skills (methods), and Open Engine (orchestration layer), rather than building it manually.
“Let your agent, like your Claude, like your Codex, help you build, right? Let it help you build open brain, which carries memories, and open skills, which gives you methods to actually get work done that are suited to you, and transport easily between Claude, between uh OpenAI's agents, between Gemini, etc., etc.”
In February, the focus was on getting agents to remember users; now (June 2026), the focus has shifted to giving agents a platform to act from on the user's behalf, with much greater power.
“And you know, in February a lot of that was about getting your agents to remember you. And now it it really is about giving your agents a platform to jump from and act from on your behalf cuz there's so much more powerful.”
The benefit of following this implementation approach is that once an agent is 'broken in', using it gets easier and easier, and now even the initial setup can be agent-assisted in a way it couldn't be in February.
“Once you start to get into the rhythm of using them, it gets easier and easier and easier and easier. And now even that first step, even getting the shoe so you can wear it, agents can help with that in a way they couldn't back in February.”
The build barrier for agent systems has dropped significantly; a few months ago it felt like a technical project (databases, SQL, configs, command line), but now Claude and Codex can walk users through the build almost by themselves.
“A few months ago, even if you believed in the open brain concept, I get that it was hard, right? The database, the setup, the SQL, the configs, the command line steps. I helped troubleshoot people through that. I've done it myself, and it felt like a technical project. It doesn't feel like a technical project anymore when Claude and Codex and tools like them can walk you through that build almost by themselves.”
Between January and June 2026, agents have gotten much better at acting and interpreting intent, with agents now often correctly understanding that 'draft' means draft and not send.
“What has changed between that moment, which was back in January, and now in June of 2026? Now, agents have got much better at acting. And intent is one of the central problems of the AI agent. We've made huge progress on connecting intent and acting just in the last 6 months. When you're chatting with the model now, you can communicate intent and it will often get it right, right? It will understand that draft response means draft, not send.”
The rapid change in agent capabilities means it should remind people that they too can do this, and they don't need to understand all the technical details (SQL databases, etc.) because agents are good enough to explain implementation in practical terms.
“I I I reminds you how quickly this is changing, and I hope that reminds you and encourages you that you too can do this. Even though this has technical primitives underneath like a SQL database, you can do it without really digging in and understanding the details of that because the agents are good enough to explain what it means in practice for you.”
Recent events (GLM 5.2 released open-source while Fable and ChatGPT 5.6 are banned) demonstrate that frontier model availability is unstable and driven by political/regulatory pressures.
“Frankly, we've seen that just in the last few days with GLM 5.2 coming out from an open source perspective while Fable and chat GPT 5.6 are all banned.”
The build barrier has dropped so far that it is worth inviting more people to build agent systems now, and agents themselves can help with the build process.
“So, jump in, the water's warm, and the ability to build is like 1/5 or less technical than it was back in February.”
Community members have already built travel planning systems, shared marketing brains, cross-tool context layers, and ticket-reading agents using the Open framework.
“People have built travel planning memory with this. People have built shared marketing brains in the community. People have built cross-tool context between Claude and ChatGPT and Kimmy. And they've built agents that can read tickets into a tracker and hand work to another agent.”
Users should pick a specific part of their life that they want to change and let an agent change it, and this is genuinely achievable now.
“Pick the part of your life that you want to change and let an agent change it for you. I I'm not kidding. You can actually do it.”
If someone knows a person with a real pain point that could be solved with an agent (or if they are a builder), they should tell that person 'this is not that hard' and encourage them that an agent can help build the solution.
“If nothing else, if you know someone who has a really big pain point and maybe it's agent shaped or you know it's agent shaped and you're a builder, go tell them that. Go say, "Hey, this is not that hard. An agent can help you build this."”
Positive examples of agent systems working are needed, and the speaker wants to see community members share their agent success stories.
“We need positive examples. In fact, if you have positive examples, if you're an open rebuild, if you you built on something else, maybe you built on G brain, I don't care. Put positive examples of why an agent made a difference in your life down here. I would like to see them. I want more people to see positive examples of agents working.”
Auto-review features in tools like Codex now prevent agents from sending emails if the user labels them as 'draft', adding a safety layer on top of intent interpretation.
“There are now things like auto review in Codex that make sure that it will not send if you say draft.”
The speaker has evolved Open Brain to include wiki-style connections inspired by Andre Karpathy's work, plus Open Skills and Open Engine, forming a complete stack for agent-to-agent task interaction.
“I've evolved the open brain system to include uh some of Andre Karpathy's work around wiki-style connections. And you You have that in open brain as part of the framework. I've added other pieces also, open skills and most recently open engine, which connects everything to an agent-to-agent task interaction framework.”
Tools like Claude and Codex can essentially build most of the Open Brain/Open Skills/Open Engine stack without requiring the user to use open-source alternatives like Open Claude.
“You may not realize this, but you already have tools like Claude and Codex that can essentially build most of this stack for you. You don't actually even need to get to open Claude, although if you have it, that's great.”
Giving agents memory and skills and a clear framework enables users to accomplish their goals, and it has never been easier to do this than right now.
“Giving your agents memory and skills and a clear framework to do stuff in in order to get to that world. And it has never, ever, ever been easier. It's never been easier.”
The speaker is waving 'the green flag' and explicitly inviting people to build now, stating that agents can help with the build process.
“So if this is you, if you're like, "Ah, I'm almost there. I want to build." This is your invitation. I am waving the green flag here. I am telling you the agents can get it done. The agents can help you build agents.”
The speaker has personally troubleshot people struggling to build these systems and realized that when they say something is 'easy', it doesn't always feel easy to users.
“But, the problem remains the build. And I have had to I have personally troubleshot people who are struggling with building these systems and I have realized that when I say something is easy, it doesn't always feel easy to you.”
Tasks that Open Brain enables include weekly planning, family logistics, protecting work blocks, feeding long-term projects, and other things the user builds the agent to do.
“and actually get your work done, and actually do things like weekly planning and family logistics, and the work blocks that you want to protect, and you feed the projects that keep getting started, or whatever it is that you are building this agent for.”
The world of AI agents is already changing, and there are many alternatives to 'open brain' emerging, which the speaker views positively.
“Now, that world is already changing under our feet. And I'm not the only one who recognized that problem. There are many alternatives to open brain out there, which is great.”
Of the 200,000 people who watched the original Open Brain tutorial, most have not actually built the system, presumably due to the technical difficulty at that time.
“I know 200,000 of you haven't built this, and I think it's worth building so you own your memory and like these guys who are trying to get approval from Washington for this and that don't own your memory.”