The argument in one line.
A three-tier memory architecture—operating manual, project folders with living memory, and long-term archive—lets you carry persistent context across Claude, ChatGPT, Cursor, and other AI platforms, eliminating context loss and making every response 10x sharper.
Read if. Skip if.
READ IF YOU ARE…
- A developer or technical founder using Claude, ChatGPT, or Cursor daily who struggles with context loss across conversations and wants a portable memory system.
- Someone building multiple concurrent AI-assisted projects who needs project-scoped context (CLAUDE.md files, memory directories) without switching platforms.
- A solo AI builder or small technical team that's already comfortable with file organization and wants to implement a three-tier memory architecture without expensive tooling.
SKIP IF…
- You're a non-technical creator or manager — this is built for developers who can manage file structures and understand prompt engineering.
- You're looking for memory solutions that work inside a single AI tool only — this system requires cross-platform portability and manual setup.
The full version, fast.
A useful memory setup for AI coding tools is a three-tier stack that lives outside any single app. Tier one is a global operating manual under 200 words covering identity, role, tone, stack, and non-negotiables, pasted into each platform's global instructions. Tier two is six to eight project folders, each with a CLAUDE.md describing goal, stack, prior decisions, and a memory map, plus a sibling memory directory for evolving notes. Tier three is long-term archive and expert knowledge in either Obsidian, for hand-edited markdown and graph views, or Pinecone, for semantic search across thousands of records. End every meaningful session with a wrap-up that writes summary, decisions, and next actions back into that archive so future prompts pull real context, not a blank slate.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:24
01 · Cold open + credibility
Promise (10x productivity, second AI brain across apps), pattern interrupt ('amnesia'), then quick credibility flex — sold last startup, builds AI businesses, drops a graph-view memory map as visual proof.
00:24 – 01:36
02 · What does great look like?
Defines the four properties of a great memory system before building anything: remembers everything, lets you edit on the fly, plugs into every platform (kills info silos), fuels every answer with context.
01:36 – 02:45
03 · Memory as input, not vault
Reframes the mental model. Every prompt silently pulls who you are, what you're shipping, what you started last month. Surfaces the failure mode of long threads ('Claude has amnesia' / 'speaking Spanish').
02:45 – 02:51
04 · Three levels framework
Names the architecture: Short (who am I), Mid (what am I doing), Long (what happened before + expert knowledge). Same answer across Claude/OpenClaw/ChatGPT.
02:51 – 04:04
05 · Layer 1 - Operating Manual (who am I)
First tier: identity, role, goals, tone, non-negotiables. Stuff that does not change weekly. ~200 words max. Lives natively in every platform's global settings; example walkthrough on Claude desktop + Anti-Gravity Customization.
04:04 – 05:06
06 · Explicit vs implicit memory + the rule
Two flavors: hard-coded instructions you write, plus the model's own learned memory growing as you converse. Lands the principle: 'the outcome of a conversation should never depend on chat history' — if it matters, write it down.
05:06 – 06:37
07 · Layer 2 - The Workshop (what am I doing)
Mid-term, project-scoped. Ask Claude to organize life/business into 6-8 categories (community, agency, startup, personal/health). One project folder per category.
06:37 – 08:04
08 · Project CLAUDE.md + memory folder
Each project gets a CLAUDE.md at root (mission, stack, decisions, memory map, references — keep under 200 lines) plus a memory/ subfolder for evolving artifacts: decisions, current-strategy, next-actions, session-summaries.
08:04 – 09:35
09 · Mutable layer + workflow
How you actually use it: open the project folder for whatever you're working on right now. Designed to be rewritten as priorities shift. Demos the same setup in Claude desktop projects, Claude Code, and Anti-Gravity — same idea, different UI.
09:35 – 11:25
10 · Layer 3 - The Arcade (long-term memory)
Third tier answers 'what happened before?' Two storage options introduced: Pinecone (vector DB for semantic search at scale) and Obsidian (markdown + graph view for hand-editable memory). Most people over-complicate this layer.
11:25 – 13:20
11 · Obsidian vs Pinecone tradeoff
Obsidian when you want to read and edit memory by hand (graphs, backlinks, strategy notes, decision frameworks). Pinecone when you want indexed semantic search across thousands of records, scale, anywhere access. Jack personally uses Pinecone.
13:20 – 14:35
12 · Conversation archive + wrap-up skill
First sub-layer of long-term: every meaningful conversation ends with a /wrap-up skill that summarizes decisions, next actions, metadata, and embeds the result into Pinecone. Indexed and searchable later by date and topic.
14:35 – 15:05
13 · Expert knowledge bases
Second sub-layer of long-term: domain-specific corpora (YouTube expertise, Hormozi business strategy). Layer 2 CLAUDE.md tells Claude which Pinecone indexes to consult for which questions — this is where the three layers interconnect.
15:05 – 15:40
14 · Building knowledge with NotebookLM
Workflow: ask Claude/ChatGPT to research a topic, auto-generate a 50-resource NotebookLM notebook, then download and vectorize into Pinecone (or keep in Obsidian).
15:40 – 16:40
15 · Firecrawl as MCP connector
Walkthrough adding Firecrawl as a Claude custom MCP connector (Connectors -> Add custom connector -> paste API key). Claims ~80% cost savings and better accuracy for agentic deep research vs default browsing.
16:40 – 16:52
16 · Recap + open loop to super-skills
Restates the three layers (who / what / before) and frames memory as only as strong as the skills supporting it. Hard cut into next-video CTA on 'super-skills that make your memory system more powerful'.
Lines worth screenshotting.
- AI memory should be treated as an import that silently loads into every prompt, not as a vault you access on request.
- A three-tier memory system (short: who am I, mid: what am I doing, long: what happened and what do I know) maps exactly to how human working memory operates.
- The outcome of any AI conversation should never depend on chat history — if it does, the memory architecture is broken.
- A global operating manual capped at 200 words forces the creator to identify what is actually permanent versus what is project-specific.
- Organizing active work into 6-8 project folders gives the AI system-scoped context without polluting the global memory with transient details.
- Long-term memory stored in Obsidian or Pinecone is retrievable on demand; memory stored only in chat threads is functionally lost after context truncation.
- NotebookLM and Firecrawl extend the memory system into external knowledge bases without requiring the AI to re-read source documents each session.
- Changing important parameters (income level, strategic focus, tech stack) is the second requirement of a good memory system — static memory becomes wrong over time.
- Information silos across Claude, ChatGPT, and Cursor are solved by a shared memory core that all platforms read from rather than each maintaining its own context.
- Native platform memory (Claude's built-in memory) improves as you talk but is not a substitute for explicit hard-coded operating instructions.
- A CLAUDE.md sibling memory directory per project is the practical implementation of mid-term memory — project-scoped, updateable, persistent.
- A portable memory system is more valuable than a platform-optimized one because it survives migrations to new tools without losing accumulated context.
Steal the three-question taxonomy.
Memory-as-infrastructure playbook
Memory is not storage — it's the plumbing that makes every prompt sharper than the last.
- Frame any AI-tooling content with the three-question collapse: Who am I? / What am I doing? / What happened before? It's screenshot-friendly and survives the platform port.
- Build the L2 Workshop as an actual file system: 6-8 project folders, CLAUDE.md at root, sibling memory/ directory with decisions / current-strategy / next-actions / session-summaries.
- Lift Jack's 'memory is an input, not a vault' line — rhetorical sibling to Joe's 'plumbing you own vs utilities you rent'.
- Use the testable rule as a sharp closer: 'the outcome should never depend on chat history' — if it matters, write it down.
- Steal the slide aesthetic: treasure-map / parchment over generic-AI gradient cards. Visual differentiation matters in this niche.
- Open-loop close to a 'super-skills' follow-up rather than a subscribe-ask — preserves the gift afterglow on tutorial content.
Terms worth knowing.
- three-tier memory system
- An architecture for AI tool memory divided into three layers: a short global profile (~200 words about the user), project-scoped CLAUDE.md files for active work, and a long-term archive for historical knowledge.
- CLAUDE.md
- A Markdown configuration file placed in a project folder that Claude Code reads as persistent instructions — used here as the mid-term (Layer 2) memory layer for project-specific context.
- second brain
- A personal knowledge management system — digital or structured — where a person externalizes thoughts, notes, and references so they can be reliably retrieved later by themselves or an AI.
- Obsidian
- A local-first note-taking and knowledge management app that stores notes as plain Markdown files on your computer — used here as the long-term archive layer for AI memory.
- Pinecone
- A vector database service used to store and search text by semantic similarity — an alternative long-term memory backend for retrieving relevant historical context for AI tools.
- NotebookLM
- Google's AI research tool that lets you upload documents and ask questions across them — mentioned here as a tool for querying long-term archived knowledge.
- Firecrawl
- A web scraping API that extracts clean text from web pages — used here to pull expert content into the long-term knowledge layer so Claude can reference it.
- vector database
- A database that stores text as mathematical embeddings and retrieves entries based on semantic similarity — enabling AI systems to find relevant memories even when the exact wording differs.
- global operating manual
- A short (~200-word) persistent document describing who the user is, their goals, preferred tools, and working style — stored in Claude's user-level memory as the foundation of all sessions.
- Anti-Gravity
- A third-party AI coding tool (alternative to Claude Code) that supports the same memory file conventions, making the three-tier memory architecture portable across AI coding platforms.
Things they pointed at.
00:14toolClaude (Anthropic)
02:10toolChatGPT
04:05toolAnti-Gravity (Google IDE)
04:06toolVS Code
08:30toolClaude Desktop projects
08:45toolCowork
10:09conceptKarpathi Obsidian setup
10:15conceptRetrieval Augmented Generation (RAG)
14:05productGlider (Jack's speech-to-text startup)
14:13channelAlex Hormozi
Lines you could clip.
00:00
“Called memory systems are a cheat code, but only if you use them properly.”
02:14
“Memory is not a vault, it's an input. Every prompt silently pulls from your stack.”
02:25
“How many times have you spoken with Claude or ChatGPT to get halfway through the conversation and it's talking complete Spanish?”
04:42
“The outcome of the conversation should never ever depend on a chat history.”
05:04
“Models forget things, they get truncated, they hallucinate. If it matters, we need to make sure we're writing it down.”
09:35
“Most people over-complicate it or don't set it up properly, meaning they get none of the benefits but all the complexity.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphorstory
00:00Called memory systems , but only if you use them properly. And in this video, I'm gonna show you exactly how to set up your own second AI brain that works across all apps with an incredibly simple setup.
00:14So you can have a memory operating system that makes you 10 times more productive and stops wasting your time. And if you don't know who I am, my name is Jack Roberts. I built and sold my last tech startup with a gazillion customers.
00:26Now I build my own AI businesses, and I just share the stuff that actually works. So if you haven't already, grab that beautiful coffee and let's dive straight in. So this is the Claude memory system, and I've spent over a year looking at memory systems.
00:40I've done a lot of research. This is the simplest one that I've actually found. So when we talk about the Claude code memory system, what do I actually mean?
00:48What does great look like? Let's define what that is before we think about what we're gonna build. Well, the first thing that a great memory system does is it remembers everything that you say.
00:56Right? Not just last time messages, not just the thread, every meaningful exchange that you've ever had with it, it remembers it and we can capture it and refer to it whenever we want to, just like here. So the first goal is it needs to remember what we've said.
01:10The second thing is that we can change the the important stuff on the fly. Right? Maybe we don't wanna focus on strategy a anymore.
01:18We wanna focus on strategy b or maybe your income's gone up and you wanna change and reflect that in the strategy. So we need the ability to change the important stuff with your role, your stack, whatever it is, we wanna be able to disregard old information.
01:31The third thing that it needs to do, it needs to plug into every platform. This is what I call information silos. It's very common AI, chatty p t for this, Claude for that.
01:41Maybe your grandmother's basement for something completely different. We don't wanna be hip and hopping. We need a central memory system, which is the universal point of truth.
01:49We call it the memory core in this system. Fourthly and finally, it needs to fuel every answer with context. What is the point of this memory system if we can't ask it questions about stuff that's happened and we can't get context rich information when we ask it things about strategy?
02:04We have a prompt, we have our three tier three level memory system, then we get the output. That is the basic idea. Memory is not a vault, it's an import.
02:12Every prompt silently pulls from your stack. Who you are, what you're shipping, what you just started last month, so response lands sharper than any fresh chat could ever hope to do. Now how many times have you spoken with Claude or ChatGPT to get halfway through the conversation and it's Spanish?
02:28And it's because amnesia. Now the best way to solve for this is to think about your memory system as solving this across three levels.
02:36Okay? We have short term memory, which is very simply, who am I? We have midterm memory, which is what am I doing?
02:42And then we have long term memory, which is what's happened before and an expert knowledge base. I'll explain exactly what I mean.
02:49Alright. Now not to get metaphysical, but the first question you have to ask yourself is who am I?
02:53And I'm not talking about glass of red wine at 4AM in the morning. I mean, does your model actually know who you are? We call this the operating manual and it's the very first level, the very first tier of our three level memory system.
03:06So think about it like this. Who am I? It's your name, your role, your current goals, the way you like your answers framed, the tools you use, tone voice, non negotiables, stuff that does not change every week, but defines every reply.
03:18So here's me for example. My name is Jack Roberts, AI YouTuber, direct me fluff, m dashes emojis, vibe is Pacheco, if I'm talking about this particular presentation. Now the idea of this is that it lives natively in every single platform and it grows as you talk.
03:32So the idea with Claude, for example, has its own internal memory system which gets better over time. And when you express preferences or ask it things, it records that for you. So as you're actually talking to them and you say, hey, remember this or I prefer x so busy, Claude will actually remember that stuff for you.
03:47But one little hack I wanna show you real quickly. So first actionable step, go on Claude. If you're coding Claude, come to the bottom left under Jack Roberts, click on obviously, yours may not say Jack Roberts, unless we got the same name.
03:57I don't know. Go up to general. And what you're looking for here is instructions for Claude.
04:01Put some stuff in here. Try to make it no more than 200 words maximum. Just top level of of how you want Claude to behave and any key information about you.
04:09Okay? Then if you're in anti gravity, for example, all you do is click on the additional options and you click on customization. This will be the same in Versus Code and various other forks.
04:19And you'll see you have global and you have workspace. You click on global and these are gonna be Gemini. Mv which are the global instructions for how these models basically behave with you.
04:27Now what's important to bear in mind here is that there's two bits of memory. One is explicit hard code stuff, and then we have the generalized knowledge it has of you, which it will natively pick up based on your conversations. So this is an important thing to understand with these memory systems is that the outcome of the conversation should never ever depend on a chat history.
04:47I should be able with this system, and you can with this system, that's why it's so cool, open new chat window and regardless of any previous message, it should always give me the best advice possible. But think of this level one as your general top of funnel memory that identifies who you are. It's important to bear in mind that models forget things, they get truncated, they hallucinate.
05:06If it matters, we need to make sure we're writing it down. Okay. So real quick, what are you doing?
05:10Well, not right now. I mean, you and are hanging out together. But I mean, what are you working on right now?
05:15Because this is the second level of our three tier system. It's the projects you're doing. They may change from time to time.
05:21Maybe you're building a startup. Maybe you've got a really cool client or several cool clients. These are the things that you'll be doing right now and that forms the second level of our system.
05:30And that's why in this second level, it's actually quite structural. So what we're going to do essentially is create six or seven separate folders for all the things that we're doing.
05:39So for example, I might have one for my community. I might have one for my agency. I might have one for my startup.
05:46And or, you know, one might be personal life and health and fitness, getting like in shape and all that kind of stuff. So what you need to do first of all is head over to Claude, or this could be your model of choice. I need to ask you this question.
05:55Hey. Based on everything that you know about me, all about conversation, all about history and projects, I want you to organize all the areas of my life and business into six to eight different categories that surmise everything that I'm doing.
06:08I'm using this to build up folders to best organize my life and my business. And once you've got that, you can basically create it. And it is surprisingly accurate, but you don't want any more than eight.
06:18Otherwise, it's way too much to manage. Now what we're gonna do based on these eight things, we'll come down to Shiva's workshop, and what we're effectively doing is creating one unique Claude dot m d file for each of those eight projects. You could do it based on clients, but again, try to keep it under eight.
06:32Alright? So we've the Claude dot m d at the root that tells Claude what the project is. The memory folder beside it stores everything that evolves, the decisions, strategies, the summaries, the next actions.
06:42You can open up in any platform and it just works. And what I've done for you as well, I've pulled together this project operating manual that you can literally copy and paste. And what I want you to do based on those eight things is come down and essentially create this per project.
06:56So this is open opens a folder. So what happens here is that Claude will open this up before it touches any of your code. So it sits as a living medium term memory that can change based on what you're working on at the moment.
07:08It's the midterm layer of our three level memory system. Okay. So what's included in this folder?
07:13Well, we need to explain what the folder is, what the goal is, why does it exist. This exists to get me 10% body fat. This exists to make me x thousand dollars.
07:21This exists to help my client do x y z. The stack, what are you building it with? Decisions you've already made.
07:28Okay? So calls already made so we don't need to relitigate them. It needs to have a memory map about where each individual memory lives and any references that are relevant.
07:36Okay? Now the template is simply this. I'll let you basically copy and paste this, but essentially fill this out for everything.
07:42Okay? And here's an example of what that might look like based on everything you're doing. Now we wanna keep it under 200 lines because this is your MD.
07:49Effectively, what this will do is be prefilled into every single conversation, so we don't want it to be too big. It needs to be updated, needs to be accurate, and it's one folder per repo. So ask all the questions, copy this into these six folders, and then fill it out.
08:01Now if you are in Claw desktop app, what you can do in Cowork, if you're using Cowork, is literally use what I would call the projects tab. So you click on projects, this is their version of that. Okay?
08:11These can be your six or seven different versions. Awesome. Now, if you're in code, again, I'd recommend code honestly because it's just everything that code does, but like a better and unlimited and it's essentially the same thing.
08:23Then in code, all you would do is effectively create these six or seven folders on your desktop. And when you open this up, you can see all the folders on the left hand side. And it's the exact same thing in anti gravity.
08:34You can have all of the folders basically, and all you do ever do is you open it up and you open up the folder that you want to chat in based on the conversations that you're having. And so the idea here is that whenever you do work, all you do is you think about which of the seven things is it about. It's about my business.
08:50Cool. When I click on business, it's about growing my LinkedIn. Then I click on LinkedIn.
08:54And effectively, do all your work inside that project for that specific thing. It has these living midterm strategy that may change. It is what we call mutable.
09:03It is subject to change and we do all the work in those seven project folders based on the things that you want to do. And the third level is probably one of the most important which is it's a long term memory. And most people I've seen either drastically over complicate it or don't set it up properly, meaning they get none of the benefits but all the complexity.
09:22So level three is . It basically answers the question, what happened before? What the hell happened before this?
09:29So let's get into a little bit detail. Now to do this, we're gonna use one of two systems. Option one is gonna be something called Pinecone, which effectively is a big database.
09:38This looks really technical. You're probably thinking, Jack, this is way out of control. In reality, they're just bookshelves, which are very, very cool.
09:46In reality, what it looks like is you can ask questions to things like this on aiojack.com. All of my YouTube videos, any content I ever create is indexed into this beautiful thing in pine cone. It is unbelievably easy to set up.
09:59And the other alternative we have is, you might have guessed it, something called Obsidian. And Obsidian itself has gained a lot of virality, especially a new system, what we call Carpathi Obsidian, was considered an alternative to what we use in Pinecone, is retrieval augmented generation.
10:15Now the idea of this is that it tracks all of the text for any topic you want to. You can amend it. Claude has access to all of it in his markdown files, which is really cool.
10:25And it understands basically everything. And the more that you basically use this, the better the system gets because it finds cross linkages between everything. So this one here is a phrase that you can see.
10:36This is an example one I have on YouTube, and you can see the relationships between each of the individual things. I know you're thinking, Jack, graphs are great, but will graphs solve all my problems?
10:46The answer to that question is yes, they will. No. They won't.
10:48Of course, they won't. But, yeah, this is the idea of Obsidian, basically. And, obviously, you can do little funny things with it, which is, like, half the fun is moving these sliders up and down, basically.
10:56But this is what they call Carpathi Obsidian, which is his alternative to a complex rag system. Now this isn't a right or wrong. You can pick either one of these two systems that you like.
11:06Let me help you make an informed decision if I may. So if you look at Obsidian, which is the node one that I showed you, you use this when you want to read and edit the memory by hand. If you wanna physically look at your memories, double click in.
11:17For example, I wanna click on this. I wanna see what this says. I wanna come down and read it.
11:21Then you may wanna go for something like Obsidian. You can create strategy note, decision frameworks, visual graphs, and backlinks that are important to you and you want zero in for it, you just want the files. Obsidian might be one for you.
11:31I'll put a full link on screen right now if you wanna double click and set up Obsidian. I'll go through the entire thing. Alternatively, you have Pinecam.
11:37If you want semantic search across thousands of records, this is way more scalable, can work anywhere. You can publish it in apps. It can be accessed from anywhere on the planet.
11:47You know, you wanna store wrap ups on every session. If it's for scale, so you wanna store longer information like books and transcripts, it is incredible for that. In other if you want a readable file, use Obsidian.
11:57If you want index as searchable summary, use Pinecone. If you're asking what I personally use, I personally use Pinecone just because I don't wanna read through everything in Obsidian, but Obsidian does really have its genuine use cases. Now there are two levels to long term memory.
12:11One is an archive of any conversation that you've ever had. Now the really cool thing is whether this is in Claude, whether this is in OpenClear, Hermes, or any system you're using, we can, if we're using Pinecone, for example, send all this information to Pinecone and we can index it. So for example, if I have a big conversation with this, all I do is I create a skill.
12:32I can come down, do forward slash wrap up, and I have a wrap up skill that takes the entire conversation and embeds it in Pinecone. Now if you wanna know how to set this this wrap up skill up, I'll put a link on screen for you so can go through my full Pinecone video that breaks down the entire system for you, my Kapathi video.
12:46And then the second level of long term memory here guys is the what I would call knowledge, Deep knowledge. So for example, I might wanna have a YouTube expertise database in my pine cone, for example.
12:57And again, you could do this with Obsidian if you like to. But the idea here is this one might be YouTube. And this could be all the information that I want on for growing on YouTube or it could be my business or whatever it is.
13:07And then what I'm doing is when I'm having conversations, your AI agent or your AI conversation can call and reach over and conversate with this knowledge base when it's relevant.
13:18And so in the project folder section section that we covered in section two, what you might say and which is where it all interconnects together. Okay. So let's say for example that one of my folders is Glider.
13:28Right? Glider is the speech to text startup that I founded, lets you yap into a computer. In the Glider folder for example, in here, I might say, look, these are the indexes in pine cone I want you to consult when I ask you strategic questions.
13:42This is my long term memory. In addition to that, if I ask you for any history, consult this index. Then in Pinecone, I might have an index over here, which could be Glider strategy or it could be business strategy or Hormozi or whatever it is, and it will call that long term memory, that expert knowledge and history to help and improve the current conversation.
14:00Now to actually build out that long term knowledge, there's two systems we can do. One is to connect this to NotebookLM. I'll put a link down below.
14:08But for example, I could say something like, hey, go and do some research on business strategy according to Alex or Mercy. Go create for me a notebook in my notebook alarm and just get me like 50 different resources as much as you possibly can and create that notebook for me. Send that straight off.
14:23Now what's really cool about this is it will actually curate and you have the power of Claude or ChatGPT 5.5 going ahead and building those notebooks for you, which is incredible. Then we can go over, grab that information, save it on our desktop, and even vectorize it into Pinecone.
14:40Or just like we did with Obsidian, have this beautiful thing over here, which I opened the graph for so you can see because it's very shiny and we like lots of graphs over here. Just use this. Now this is my, you know, Hormozi business strategy thing that I can ask questions to.
14:52Again, I'll put a link on screen for the Notebook LM deep dive if you wanna double click and learn more about that strategy. The other one that I absolutely love doing is Firecrawl. So for example, I could say something like this.
15:03Hey. I want you to use my Firecrawl integration and go ahead and do some deep research for me on Hormozi's best five practices for scaling a business past $10,000 a month.
15:15Okay. Do that. Send that one off.
15:17Of course, we can connect Firecrawl to the connectors on plus. Click on connectors over here. You can see I've got Firecrawl installed.
15:22And all you literally do is come over, you grab the API key, you can come back over to the connectors, then what you would do is click on the plus, click on add custom connector, and then right here, you just type in Firecrawl, and then remote MCP server, just enter in this right here. Where I've got API key, you just enter your Firecrawl.
15:37And then basically, what that will then be able to do is use the FireCall agent, and this can save you like 80% of the cost, and it's just way more efficient in getting accurate data for agents. So what this basically means is we have an archive of every conversation. So every meaningful conversation ends with a wrap up.
15:52That summary decisions, next actions, made the metadata, the summary lands in the archive. So if I ever say to, hey, what was that thing we discussed about this really important event last January?
16:02We've got it. And when we do index it, we have the date and we can set all those different filters in Pinecone or locally on a computer if we're using Obsidian, which is really cool. That's the exact process we went through.
16:13Then we covered the expert knowledge that we've embedded. We've used Firecrawl to go get some deep research. We can use NotebookLM to build beautiful, detailed, impressive notebooks on any topic we want to using the world's strongest AI research bottle, which is freaking incredible.
16:28And we can pull that down anytime we want to in our memory system. So we have the three layers. We have short term, midterm, and long term.
16:34Level one is who we are. Level two is what we're doing, and level three is everything that's happened before and that perfect knowledge. Now your memory system is only strong as the skills that support us.
16:44So what we need to do next is learn what I call super skills that can make your memory system even more powerful, and we're gonna learn that by watching this video right here.
The Hook
The bait, then the rug-pull.
The cold open does double duty: dangle a fast outcome ('10x more productive') AND name the cost of inaction ('amnesia' — the moment your model starts speaking Spanish halfway through a thread). Inside sixty seconds you know the promise, the problem, and the format of the answer.
Named ideas worth stealing.
02:45model
Three-Level Memory System
- L1 Short / Operating Manual - Who am I?
- L2 Mid / The Workshop - What am I doing?
- L3 Long / The Arcade - What happened before + expert knowledge
Stratified memory across timescales and scopes. Each tier answers a different question and lives in a different location (global settings / project folder / vector store).
Steal forJoe's own AI-memory taxonomy across ~/.claude/CLAUDE.md, joe-profile.md, per-project CLAUDE.md, and MEMORY.md — name the system, not just the files.
00:24list
Four properties of a great memory system
- Remembers everything you said
- Lets you change the important stuff on the fly
- Plugs into every platform (kills info silos)
- Fuels every answer with context
Spec sheet for evaluating any memory architecture. Used as the pre-build checklist before showing the implementation.
Steal forTemplate for evaluating any infrastructure pitch in 'Own your stack' content — define great first, then build to spec.
06:37concept
Project Operating Manual template
- What is the folder / what is the goal / why does it exist
- The stack (what you're building it with)
- Decisions already made (so we don't relitigate)
- Memory map - where each memory lives
- References
Per-project CLAUDE.md skeleton. Under 200 lines because it gets prepended to every conversation in that scope.
Steal forJoe's MCN / ModBoard / Clip Lab / JoeFlow folders could each get this exact skeleton.
02:12concept
Memory is an input, not a vault
Reframes memory as plumbing into every prompt rather than a passive store you look things up in. Every prompt silently pulls who you are, what you're shipping, what you started last month.
Steal forDirect rhetorical match for Joe's 'plumbing you own vs utilities you rent' frame — AI memory as plumbing.
04:42concept
Outcomes-should-never-depend-on-history rule
Stress test for your memory system: open a new chat with zero context — does it still give the best advice? If not, the implicit/chat-history layer is doing work that should be explicit.
Steal forSharp testable assertion for any AI-tooling content Joe ships.
How they asked for the click.
VERBAL ASK
16:40next-video
“Your memory system is only strong as the skills that support us. So what we need to do next is learn what I call super skills that can make your memory system even more powerful, and we're gonna learn that by watching this video right here.”
Open-loop close — no subscribe ask, no link, no product. Pure retention CTA pointing to the next video. Clean for tutorial format because it preserves the 'I just gave you something useful' afterglow.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
