---
title: "Karpathy's LLM Wiki concept, in one read"
url: https://memory.wiki/T7ZGdpOm
updated: 2026-05-14T17:52:48.410Z
hub: https://memory.wiki/hub/demo
bundle_count: 1
concept_count: 12
source: "memory.wiki"
---
# Karpathy's LLM Wiki concept, in one read

## Source

Andrej Karpathy, Twitter thread (2024). Topic: a personal LLM-readable wiki — one place a person writes their knowledge, the LLM reads it instead of building memory by inference.

## Core argument

> The most reliable AI memory is the one the human authored. Anything inferred from a chat transcript is lossy; anything written deliberately is durable.

This frames "memory" as a *curation problem*, not a *retrieval problem*. Most existing AI memory systems (Mem.ai, ChatGPT memory beta) treat it as the latter.

## Where Karpathy's vision stops

Single unified wiki. One person, one wiki. The structure is whatever the user makes inside that wiki.

## Where mdfy goes further

Three composable scopes instead of one: doc, bundle, hub. The same URL primitive scales from a one-note share to a project context to a full knowledge graph. Karpathy's single-wiki model can't deliver per-project context without folder discipline; mdfy's bundle model is *built* for it.

But mdfy and Karpathy's vision share the load-bearing claim: **the human stays the author**. The AI reads what was written; it doesn't try to infer memory from chat.

## Quote worth keeping

> The wiki you wrote yesterday is the context the AI gets today. The AI shouldn't be reconstructing your beliefs from session transcripts.

That's the philosophy. mdfy is one shape of the implementation.


---

## Concepts in this document
- **mdfy** _(entity)_
  A tool that stores project context and decision history, integrated into Cursor via custom rules.
- **Claude** _(entity)_
  Anthropic's AI model cited as an example of vendor lock-in through projects and memory features.
- **ChatGPT** _(entity)_
  Example of an AI provider whose memory feature is intentionally confined to its own product.
- **Hub-shaped memory** _(concept)_
  The chosen architecture where users author structured notes as URL-addressable resources that the AI reads, preserving human control over memory curation.
- **Knowledge Management** _(tag)_
  Domain of organizing, storing, and retrieving information for human and AI use.
- **Vector recall** _(concept)_
  Memory pattern using embeddings and cosine similarity retrieval; trades precision for breadth but risks memory drift beyond user control.
- **URL-addressable knowledge** _(concept)_
  Design principle enabling AI systems to read structured user-authored content from persistent, addressable locations outside the AI interface.
- **Episodic snapshots** _(concept)_
  Memory pattern storing full conversation transcripts indexed by date and topic; trades verbosity for fidelity but also risks uncontrolled memory drift.
- **Composable scopes** _(concept)_
  Three-tier architecture (doc, bundle, hub) that scales from single notes to full knowledge graphs.
- **Andrej Karpathy** _(entity)_
  The originator of the LLM wiki concept presented in a 2024 Twitter thread.
- **AI Memory** _(tag)_
  Broad category encompassing how AI systems maintain and access knowledge across sessions.
- **Human-authored memory** _(concept)_
  The core principle that durable AI memory comes from deliberate human writing, not lossy inference from chat transcripts.

## Concept relations (within this doc's concepts)
- **Vector recall** implemented by **ChatGPT**
- **Episodic snapshots** implemented by **Claude**
- **Hub-shaped memory** implemented by **mdfy**
- **Composable scopes** enables pattern **URL-addressable knowledge**
- **Hub-shaped memory** implements principle **Human-authored memory**
- **Composable scopes** extends architecture **Hub-shaped memory**
- **URL-addressable knowledge** enables approach **Hub-shaped memory**
- **mdfy** implements pattern **Hub-shaped memory**
- **mdfy** provides architecture **Composable scopes**
- **mdfy** integrates with **Claude**
- **mdfy** solves problem for **ChatGPT**
- **Hub-shaped memory** depends on **URL-addressable knowledge**

## Bundles containing this document
- [AI Memory Research](https://memory.wiki/b/wa-K_2rF)
  > Captured conversations + external reading on how AI memory architectures actually work. Reading order: the Claude conversation lays out the three patterns, the Karpathy summary names the philosophical

_Hub canonical:_ https://memory.wiki/hub/demo
_Concept digest:_ https://memory.wiki/raw/hub/demo?digest=1&compact=1