AI Vibe coders have another reason to thank Andrzej Karpathy, the coiner of the term.
The former director of AI at Tesla and co-founder of OpenAI, who now runs his own independent AI project, described in a recent post on "LLM Knowledge Base" The approach he is using to manage different topics of research interest.
By creating a persistent, LLM-preserved record of his projects, Karpathy is solving the original frustration "stateless" AI Development: The Dangerous Reference-Limit Reset.
As anyone who has Vibe coded can attest, exceeding usage limits or terminating a session often feels like a lobotomy for your project. You are forced to spend valuable tokens (and time) in the hope of reconstructing the context for the AI. "Memory" The architectural details you just established.
Carpathy offers something simpler and more loosely, messily elegant than the typical enterprise solution of vector databases and RAG pipelines.
Instead, he outlines a system where the LLM himself works as a full-time "research librarian"- Dynamically compiling, lining up and interlinking Markdown (.md) files, the most LLM-friendly and compact data format.
by diverting a large part of it "token throughput" In the manipulation of structured knowledge rather than boilerplate code, Karpathy has created a blueprint for the next phase "second brain"- which is self-healing, auditable and completely human-readable.
Beyond RAG
For the past three years, the dominant paradigm has been to provide LLMs with access to proprietary data Recovery-Augmented Generation (RAG).
In a standard RAG setup, documents are shredded arbitrarily "pieces," Converted into mathematical vectors (embeddings), and stored in a special database.
When a user asks a question, the system executes "similarity search" Finding the most relevant parts and putting them into LLM.Karpathy’s approach, which he calls LLM Knowledge BaseNegates the complexity of vector databases for medium-sized datasets.
Instead, it depends on the LLM’s increasing ability to reason over structured text.
System architecture, as seen by X user @himanshu In part the extensive responses to Karpathy’s post, works in three distinct steps:
- Data Ingestion: The raw material – research papers, GitHub repositories, datasets, and web articles – is dumped into a
raw/Directory. Carpathian uses obsidian web clipper To convert web content to Markdown (.md) files, ensuring that the images are also stored locally so that the LLM can reference them through vision capabilities. -
Compilation Steps: This is original innovation. Instead of just indexing files, LLM "Compilation" Them. It reads raw data and writes a structured wiki. This includes preparing summaries, identifying key concepts, writing encyclopedia-style articles, and importantly composing posterior between related ideas.
-
Active Maintenance (Linting): The system is not stable. Karpathi describes running "health checkup" Or "linting" Passes where the LLM scans the wiki for anomalies, missing data, or new connections. As a member of the community Charlie Wargnier saw, "It acts as a living AI knowledge base that actually heals itself."
assuming markdown files "source of truth," Carpathians avoid it "black box" Vector embedding problem. Every claim made by AI can be traced back to a specific basis .md A file that a human can read, edit, or delete.
Implications for Enterprise
While the Karpathy establishment is currently described as a "hacky collection of scripts," The implications for enterprise are immediate.
As entrepreneur Vamshi Reddy (@tammireddy) said in response to the announcement: "Every business has a directory. No one ever compiled it. That’s the product."
Karpathy agreed, suggesting that this method represents a "incredible new product" Social class.
Currently most companies "Drown" In unstructured data – Slack logs, internal wikis, and PDF reports that no one has time to synthesize.
A "karpathi style" The enterprise layer won’t just search these documents; This will be actively authored "Company Bible" Which updates in real time.
As AI teacher and newsletter writer Ole Lehmann puts it on X: "I think whoever packages this for normal people is onto something big. An app that syncs with the tools you already use, your bookmarks, your read-it-later apps, your podcast apps, your saved threads."
Eugen Alpezza, co-founder and CEO of AI enterprise agent builder and orchestration startup Adra, said in an X post that: "The jump from personal research wiki to enterprise operations is where it gets brutal. Thousands of employees, millions of records, tribal knowledge that conflicts between teams. Actually, there is room for a new product and we are building it in the enterprise."
As the community discovers it "karpathi pattern," The focus is already shifting from individual research to multi-agent orchestration.
A recent architectural breakdown by @jumperz, founder of AI agent creation platform Secondmate, shows this evolution through a "herd encyclopedia" Which scales the Wiki workflow to a 10-agent system managed via OpenClaw.
The main challenge of multi-agent swarm – where a hallucination can get complicated "infected" Collective memory – is addressed here by a dedicated "Quality Door."
Using the Hermes model (trained by Nous Research for structured assessment) as an independent observer, each draft article is scored and validated before being promoted. "stay" Wiki.
This system creates a "compound loop": Agents dump the raw outputs, the compiler organizes them, Hermes verifies the truth, and verified briefings are sent back to the agents at the beginning of each session. It makes sure the herd never comes "wakes up empty," Instead each task begins with a filtered, high-integrity briefing of everything the collective has learned
Scaling and performance
A common criticism of non-vector approaches is scalability. However, Karpathy says that at a scale of ~100 articles and ~400,000 words, LLM’s ability to navigate through summary and index files is more than adequate.
For a departmental wiki or personal research project, "fancy melody" infrastructure often introduces greater latency and "recovery noise" This solves it.
Tech podcaster Lex Fridman (@lexfridman) confirmed that he uses a similar setup, adding a layer of dynamic visualization:
"I often have it generate dynamic HTML (with JS) that allows me to sort/filter the data and interactively tinker with the visualizations. Another useful thing is that I have the system generate a temporally focused mini-knowledge-base… which I load into the LLM for voice-mode interactions on long runs of 7-10 miles."
it "ephemeral wiki" The concept suggests a future where users simply don’t "to talk" With AI; They set up a team of agents to create a custom research environment for a specific task, which is finished once the report is written.
Licensing and ‘file-over-app’ philosophy
Technically, Karpathy’s functionality is built on an open standard (Markdown), but it’s viewed through a proprietary-but-extensible lens (note-taking and file organization app Obsidian).
- Markdown (.md): By choosing Markdown, Karpathy ensures that its knowledge base is not limited to a specific vendor. It is future-proof; If Obsidian disappears, the files remain readable by any text editor.
-
Obsidian: While Obsidian is a proprietary application, it "local-first" The philosophy and EULA (which allows free personal use and requires a license for commercial use) align with the developer’s desire for data sovereignty.
-
"vibe-coded" tool: The search engines and CLI tools that Carpathy mentioned are custom scripts – possibly Python-based – that bridge the gap between the LLM and the local file system.
it "file-over-app" The philosophy is a direct challenge to SaaS-heavy models like Notion or Google Docs. In the Carpathy model, the user owns the data, and the AI is merely a highly sophisticated editor. "seizure" Files to work with.
Librarian vs Search Engine
The AI community has responded with a mix of technical validation and "vibe-coding" Excitement. The debate focuses on whether the industry has over-indexed on vector DBs for problems that are fundamentally about structure, not just similarity.
Cloud-using welder Jason Paul Michaels (@SpaceWelder314) echoed the sentiment that simpler tools are often more robust:
"No vector database. No embeddings… just markdown, FTS5, and grep… every bug fixed… gets indexed. Knowledge is compound."
However, the most significant praise came from Obsidian co-creator Steph Ango (@Kepano), who highlighted a concept "Contamination Mitigation."
He suggested that users should keep their views private "arch" Clear and let the agents play "dirty safe," Fetching useful artifacts only when agent-facing workflows have distilled them.
Which solution is right for your enterprise vibe coding project?
| Speciality |
vector db/rag |
Karpathy’s Markdown Wiki |
|
date format |
opaque vector (mathematics) |
human-readable markdown |
|
logic |
Semantic similarity (nearest neighbor) |
Clear connections (backlinks/indexes) |
|
Audit |
low (black box) |
High (direct detection capability) |
|
compounding |
static (requires re-indexing) |
Active (self-healing through linting) |
|
ideal scale |
millions of documents |
100 – 10,000 high-signal documents |
"vector db" The approach is like a huge, disorganized warehouse with a very fast forklift driver. You can find something, but you don’t know why it’s there or what its relation is to the pallet next to it. Karpathy’s "markdown wiki" Like a curated library with a head librarian who is constantly writing new books to explain the old ones.
next step
Karpathy’s final discovery points to the ultimate destination of this data: synthetic data generation and fine-tuning.
As the wiki grows and there is more data "Pure" Through continuous LLM lining up, it becomes the perfect training set.
Instead of LLM just reading wiki in it "reference window," The user can ultimately improve a smaller, more efficient model on the wiki itself. This will allow to do LLM "Know" The researcher’s personal knowledge base lies in his or her own weight, essentially turning a personal research project into a custom, private intelligence.
The bottom line: Carpathy hasn’t just shared a script; He shared a philosophy. By treating the LLM as an active agent that maintains its own memory, it has circumvented its limitations. "one shot" AI interaction.
For the individual researcher, this means the end "Forgotten bookmark."
For the enterprise, this means the transition from a "RAW/DATA LAKE" to one "Compiled knowledge wealth." As Karpathy himself summarized: "You rarely write or edit the wiki manually; This is the domain of LLM." We are entering the era of autonomous storage.
<a href