What this covers

Chase walks through a seven-level ladder for giving Claude Code a usable memory, starting with its built-in auto-memory and ending with agentic RAG. The framing is broader than the title implies: RAG is only one rung among several, and the real subject is how to deal with AI memory problems without over-building. The opening levels cover how Claude Code writes its own markdown notes into a dot-Claude memory folder, why most people never clear a session for fear of losing past work, and how that habit feeds Context Rot. He cites a drop from 92% to 78% performance as the window fills from 256k to 1 million tokens, a figure the extraction marks as contested.

The middle levels get practical. He argues that a Claude.md injected into every prompt can add noise rather than signal, and he proposes a multi-file layout (project, requirements, roadmap, state) that gives Claude a navigable hierarchy. Level 4 is Obsidian, with a nod to Andrej Karpathy's widely viewed LLM knowledge-base approach. He calls it the 80-90% solution for solo operators, while admitting there is no clean document-count threshold at which it breaks, since token cost and speed matter alongside accuracy.

Levels 5 through 7 are true RAG. He explains naive rag (chunking, embedding into a vector database, retrieving similar chunks), then its weaknesses: lost context between chunks, trouble relating documents, and roughly 25% accuracy without reranking or graphs. graph rag with Light Rag is presented as the minimum worthwhile infrastructure for relationship questions. Level 7 adds multimodal retrieval, data pipelines, versioning and access control, and Gemini Embedding 2's direct video embedding. His recurring advice is to start simple, since moving from complex to simple is costly, and to distrust hype.

Sharpest takeaway

Claude Code and AI systems need structured memory architectures beyond native auto-memory, progressing through seven levels from simple markdown files to sophisticated agentic RAG systems, with most users finding Obsidian-based solutions sufficient for their needs.

  • Auto-memory and bloated context windows create context rot that degrades AI performance over time
  • Most users stay at level one due to psychological fear of forgetting conversations, but this creates efficiency problems
  • Obsidian with proper file structure can handle thousands of documents without traditional RAG, making it the practical sweet spot for solo operators

The claims · ranked29 claims · weighted by value

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.

0.74

Unorganized document systems (like billions of files strewn across a factory floor without filing system) make it nearly impossible for AI systems to find information, highlighting that organization architecture matters more than raw storage capacity or sophisticated retrieval algorithms.

factualhigh valueestablishednovelty 1/4durability 4/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“It's the equivalent of having 10 million files strewn across the factory floor. I mean like, 'Well, Claude code, find it.' Like no, you actually just need a filing cabinet. Like Claude code's actually pretty smart.”

0.70

Claude.md files, when referenced in every prompt, can actually reduce the effectiveness of large language models at large according to studies evaluating agent.md files, because injecting the same context into every prompt creates noise rather than signal if that context isn't universally relevant.

causalhigh valuecontestednovelty 3/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“When we look at studies like this one evaluating agents.md and you can swap agents.md for Claude.md, they found in the study that these sort of files can actually reduce the effectiveness of large language models at large. And why is that? Well, it's because the thing that makes it so good, the fact that it's injected into basically every prompt, is what also can make it so bad.”

0.68

Context rot is a phenomenon where AI system effectiveness degrades as more tokens fill the context window in the same session, with Claude demonstrating a drop from 92% to 78% performance as the context window fills from 256k to 1 million tokens.

causalhigh valuecontestednovelty 2/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“context rot is the phenomena that the more I use an AI system within its same session, within its same chat, and I fill up that context window, the worse it gets. You can see that right here. Claude Code 1 million context window at 256k tokens, aka I've only filled up about a quarter of its context window, we're at 92%. By the end of it, 78.”

0.65

Key limitations of Level 6 (Graph RAG) include inability to handle non-text documents like scannable PDFs, videos, and images, and limited multimodal retrieval—these problems require Level 7 solutions.

factualhigh valueestablishednovelty 1/4durability 3/4· Chase (Channel Creator)

“When we talk about Light Rag, this is just text. What if I have scannable PDFs? What if I have videos? What if I have images? We don't live in a world where all your documents are just going to be Google Docs. And so, what do we do in those instances?”

0.65

In naive RAG systems, chunking documents into pieces and retrieving individual chunks can cause context loss when chunks reference information in other chunks not retrieved, and creates difficulty in understanding relationships between different chunks or documents.

causalhigh valueestablishednovelty 1/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“What if chunk number three is referencing something in chunk number one, and then our vector situation when we pull the chunks, what if it doesn't get the right one? What if it doesn't get that other chunk that's required as context to even make sense of what number three says?”

0.59

The trap at Level 7 is trying to force yourself into this level of sophistication when it's not needed, and most users are fine with Obsidian and don't need RAG in general, let alone Level 7 agentic RAG.

normativehigh valuespeaker onlynovelty 2/4durability 4/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“the trap, honestly, is trying to force yourself into this level and this sort of sophistication when it's just not needed. To be honest, after all this, most of you're fine with Obsidian. This is more than enough. You don't need graph rag. You really don't need rag in general. And if it's not obvious that you need level seven, and certainly if you haven't already tried the Obsidian route, you don't need to be here.”

0.59

Always start with simple solutions like Obsidian and scale up to sophisticated RAG systems only if necessary, because moving from simple to complex is easy but moving from a locked-in complex system to simple is difficult and costly.

normativehigh valuespeaker onlynovelty 2/4durability 4/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“And so I always suggest to everybody, clients included, like let's just start with Obsidian and see how far we can scale this. And eventually, if we do hit a wall, you can always transition to more robust rag systems. So why not try the simple option? If it works, great, it's free, cost me no money. Versus like, let's try to knock out this rag system which can be kind of difficult to put into production”

0.55

In vector databases, vectors are points in multi-dimensional space (hundreds to thousands of dimensions) represented by series of numbers, where placement depends on semantic meaning so related concepts cluster together (e.g., fruits vs boats in different regions of the space).

definitionhigh valueestablishednovelty 0/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“in a vector database, it's not two-dimensional columns and rows. It's actually hundreds if not thousands of dimensions. But for the purposes of today, just think of a three-dimensional graph like you see here. And the vectors are just points in that graph. And each point is represented by a series of numbers.”

0.55

Naive RAG is the most basic type of RAG system consisting of three parts: embedding stage, vector database, and retrieval with LLM, where documents are chunked, embedded into vectors placed in multi-dimensional space based on semantic meaning, then retrieved by embedding the query and finding similar vectors.

definitionhigh valueestablishednovelty 0/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“you can kind of think of rag systems being broken out into three parts. On the left-hand side, we have the embedding stage. We then have the vector database. And then we have the actual retrieval going on with the large language model.”

0.52

Level 7 agentic RAG requires attention to data pipelines and architecture for data ingestion/syncing, not just retrieval, because real systems need versioning, deduplication, access control, and update mechanisms that production-level RAG systems must handle.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“level seven is also where we need to start paying attention to our architecture and pipelines when it comes to the data going in and out of our rag system. It's not enough to just get data in here. How is the data getting there? How is the data getting there in the context of a team? How is data getting out of there? Like, what if some of the information here has changed in a particular document? What if somebody edits it? How does it get updated? What if we add duplicates?”

0.52

The right RAG approach depends on what questions you're asking: if you're treating knowledge base as a giant rulebook needing specific items, Obsidian or naive RAG suffices; if you need to understand relationships between X and Y in separate documents, you need graph RAG.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“what kind of questions are you actually asking about your system? If you're just asking, you know, essentially treating your knowledge base as a giant rulebook and you just need specific things from that knowledge system brought up, well then Obsidian's probably enough. But if we need to know about relationships, if we need to know about how X interacts with Y and they're two separate documents, they never even really mention each other, and it's not something I can just stick inside the context directly because I have thousands of said documents. Well, that is where when you're going to need rag, and that's when you're going to need something more sophisticated than basic vector rag.”

0.52

Most people use Claude Code inefficiently by never clearing conversations or resetting sessions, instead continuously filling up the context window because they fear the AI will forget past work when they exit, creating a cycle where token usage increases dramatically and performance deteriorates.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“what is everybody's answer to the chat window not being able to remember anything? Well, the answer is you just keep that conversation going forever because you don't want to get to a scenario where you have to exit out and it forgets everything. This is a fear that is born here inside of these chat windows, beginning with ChatGPT and same thing with Claude's web app.”

0.52

Levels 5, 6, and 7 involve true RAG structures that are overkill for most people; Obsidian at Level 4 is the 80-90% solution that suffices for most solo operators because it's free, requires minimal overhead, and solves the core problem of connecting Claude Code to many documents with accurate retrieval.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“Obsidian is that 80% solution that in reality is like a 99% solution for most people because it's free, there's basically no overhead, and it does the job for the solo operator. And when I say it does the job for the solo operator, I mean it solves the problem of having Claude code connected to a bunch of different documents, a bunch of different markdown files, and being able to get accurate, timely information from it”

0.50

There is no clear threshold number for when Obsidian systems stop working because all documents are different, with scale concerns being more complex than just document count—including token costs and execution speed rather than just accuracy.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“The answer is we don't really know and there isn't a clear number because all your documents are also different. And in terms of hitting a wall, it isn't just as simple as, well, Claude code's giving us the wrong answers. It has too many files in Obsidian system. How much is it costing you in terms of tokens now that we've added so many files and how quickly is it doing it?”

0.50

The multi-file architecture approach with specialized files (project, requirements, roadmap, state) fights against context rot by avoiding the injection of all files into every prompt like Claude.md does, and creates a clear navigation hierarchy for Claude Code to follow.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“By breaking up memory and context and conventions in this sort of system, we're fighting against the idea of context rot and the idea brought up in that study, which is injecting these files into every prompt all the time like we do in Claude.md. It's actually counterintuitive. It doesn't help us get better outputs.”

0.48

The overall roadmap through the seven levels shows where users stand in the fight between AI systems and memory, helping them understand both what they can do to improve and what they actually need versus what hype suggests they should want.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“this video is really about deconstructing that problem of Claude Code and really AI systems in general and memory. And even more importantly, this video is about giving you a roadmap that shows you where you stand in this fight between AI systems and memory and what you can do to get to the next level.”

0.48

Understanding naive RAG fundamentals is essential even if not implementing it, because it informs decisions about document structure and system architecture when building sophisticated RAG approaches, and helps identify when naive RAG is actually sufficient.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“level five isn't about implementing these sort of naive rag systems. It's about understanding how they work so that you when it comes time to implement something more sophisticated, you actually understand what's going on.”

0.45

Simple naive RAG systems without sophisticated reranking or graph approaches often achieve only ~25% accuracy, making them barely better than random guessing, and people unfamiliar with RAG limitations risk being "hoodwinked or confused or in some cases scammed" into buying ineffective systems.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“it's going to suck to the tune of like, 'Oh, the actual effectiveness of this is like 25% of the time you get something right.' Like you're almost better guessing.”

0.43

Graph RAG (Level 6) is the 'lowest level of infrastructure you need to create' if you're actually going to use RAG, using tools like Light RAG which is completely open-source and lightweight.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Chase (Channel Creator)

“if you are going to use rag, this is sort of the lowest level of infrastructure you need to create. This is using Light Rag, which is a completely open-source tool.”

0.43

Claude Code automatically creates markdown files in its memory system that use its own intuition to decide what information is important about a project, storing these in a dot-Claude folder under a memory subdirectory.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“the auto memory system, which is automatically enabled when you use Claude Code, essentially allows Claude Code to create markdown files on its own that sort of list out things it thinks are important about you and that particular project. And this is purely based off of its own intuition based on your conversations. And I can see these memory files it's created again, it does this on its own. If you go into your dot Claude folder, you go into projects, you will see a folder there that is called memory.”

0.39

Andre Karpathy's approach to building LLM knowledge bases on top of Obsidian has received almost 20 million views, indicating significant interest in this architecture and validating it as a benchmark for practical knowledge base design.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Chase Lean

“When you have people like Andre Karpathy talking about these LLM knowledge bases they've created which are built for all intents and purposes on an Obsidian foundation and it's getting almost 20 million views, we should probably listen and see how this is actually operating.”

0.39

Gemini Embedding 2, released in March 2026, enables direct embedding of video content (not just transcripts) into vector databases, representing a significant new capability for multimodal RAG systems.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Chase Lean

“We also have new releases like Gemini embedding, too, which just came out in March, which allows us to actually embed videos into our vector database, videos itself.”

0.39

The video title 'Seven Levels of Claude Code and RAG' initially suggests RAG focus but the actual thesis is about deconstructing AI memory problems broadly, with RAG being just one of several approaches at different maturity levels.

factualhigh valuespeaker onlynovelty 0/4durability 3/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“And while this video is titled the seven levels of Claude Code and rag, what this video is really about is deconstructing that problem of Claude Code and really AI systems in general and memory.”

0.36

Comparison studies showing 1,200x differences between textual RAG and textual LLMs are outdated (2025 or earlier), used non-Claude models, and the gap has likely shrunk significantly with modern models, making the exact performance ratio uncertain.

factualhigh valuespeaker onlynovelty 1/4durability 1/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“This was done in 2025. This was not done with Claude code. These models have changed significantly since then. These are just straight-up LLMs. This isn't a coding artist. Etcetera, etc., etc. However, we were talking a 1,200x difference. So, when we're evaluating... the answer is you have to experiment and you need to try both and see what works. Because this is frankly out of date. Totally. Like 2025, older models. The difference between rag and textual LLMs is not 1,200 times. But how much has that gap shrunk?”

0.35

Graph RAG (Level 6) uses Light Rag and graph-based approaches to extract entities and relationships from documents, creating a network where everything is interconnected and queries can understand relationship patterns, showing 100%+ performance improvements over naive RAG in benchmarks.

factualhigh valuespeaker onlynovelty 0/4durability 2/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“when we compare naive rag to Light Rag across the board, we get jumps of often times more than 100%, right? 31.6 versus 68.4, 24 versus 76, 24 versus 75, on and on and on.”

0.35

Light Rag is the lightest-weight graph RAG system available, with Microsoft's Graph RAG being more robust but more complex and expensive, and according to Light Rag benchmarks, Light Rag beats Graph RAG (though these are Light Rag's own numbers and should be taken with skepticism).

factualhigh valuespeaker onlynovelty 0/4durability 2/4· Unidentified Speaker — The 7 Levels of Claude Code & RAG [kQu5pWKS8GA]

“Light Rag is the lightest weight graph rag system out there that I know of. There are some very robust versions including graph rag itself from Microsoft. It's a graph It's literally called graph rag. But when we compare naive rag to Light Rag across the board, we get jumps of often times more than 100%, right? And that being said, according to Light Rag, it actually holds its own and beats out graph rag itself, but hey, these are Light Rag's numbers, so take it with a grain of salt.”

0.34

Re-rankers are a technique used to improve naive RAG by taking the set of retrieved vectors and doing a second pass with a large language model to rank them by relevance, adding sophistication that addresses some (but not all) limitations of basic vector similarity retrieval.

definitionestablishednovelty 0/4durability 4/4· Chase Lean

“over the years, we've come up with some ways to alleviate these issues, things like re-rankers or ranking systems that take a look at all of the vectors we grab and essentially then do another pass on them with a large language model to rank them in terms of their relevance.”

0.17

RAG Anything is a tool that enables importing and ingesting images, tables, and non-text documents (scannable PDFs) into structured knowledge graphs like Light RAG, addressing the limitation of text-only RAG systems.

definitionspeaker onlynovelty 0/4durability 2/4· Chase Lean

“things like rag anything, which allow us to import images and non-text documents. Again, think scannable PDFs into structures like the light rag knowledge graph you saw here.”

0.13

The Claude Code Masterclass inside Chase AI Plus teaches production-level RAG implementation and real-world deployment scenarios where practitioners work in teams or sell RAG systems to clients.

factualspeaker onlynovelty 0/4durability 1/4· Chase Lean

“the Claude Code Masterclass and is the number one way to go from zero to AI dev, especially if you don't come from a technical background. And this Masterclass is a little bit different because we focus on a number of different use cases to learn how to use Claude Code. One of those is something like production level rag, how to build the rag systems you're going to see in this video in a real-life scenario and actually use it as a member of a team or sell it to a client.”