Karpathy's LLM Wiki vs. Waykee Cortex

The Problem Both Systems Solve

In April 2026, Andrej Karpathy — former AI Director at Tesla and co-founder of OpenAI — published llm-wiki.md, a pattern for building personal knowledge bases using LLMs. His key insight resonated with the entire AI community:

Both Karpathy's LLM Wiki and Waykee Cortex reject the dominant paradigm of treating AI as a search engine (ask a question → retrieve fragments → generate answer → forget everything). Both believe that knowledge should compound, not be re-discovered on every query.

But they take fundamentally different approaches to solving this problem. Understanding the difference helps teams choose the right tool — or combine both.

Architecture: Three Layers vs. Three Levels

Both systems are organized in three layers, but the nature of each layer differs significantly.

Layer 1: Where Knowledge Lives

Karpathy's raw/ directory is a flat collection of immutable source documents — PDFs, articles, datasets, images. The human curates what goes in; the LLM never modifies these files. They are the "source of truth."

Waykee's Knowledge hierarchy organizes documentation in a structured tree: System → Module → Screen/Program. Each node has rich metadata (GitHub repos, ports, file paths), HTML documentation for humans, and plain-text BotContext optimized for AI consumption. Knowledge is not immutable — it's a living record that gets updated when the system it describes changes.

Layer 2: How Knowledge Gets Delivered

Karpathy's Wiki layer is the crown jewel of his system. The LLM reads raw sources and compiles them into interlinked markdown pages — summaries, entity profiles, comparison tables, synthesis articles. A central index.md catalogs every page. When you ingest a new source, the LLM touches 10–15 wiki pages, updating cross-references and resolving contradictions.

Waykee's BotContext engine dynamically assembles context on demand. When an agent calls GetBotContext for any waykee (task, screen, module), the engine walks up the hierarchy, collecting documentation and metadata from each level. A single API call returns the complete context chain: the screen's docs, its parent module's docs (with GitHub repo and port), the parent system's docs, plus conversation history and instruction packs.

Layer 3: How Behavior Gets Configured

Karpathy's Schema (CLAUDE.md or AGENTS.md) is a configuration file that defines wiki structure, conventions, and workflows. It co-evolves with the human over time — a single document that shapes how the LLM interacts with the wiki.

Waykee's Instruction Packs are modular, reusable instruction sets that can be independently assigned to different agents. A "Developer Guidelines" pack, a "UTC DateTime Rules" pack, and a "Deployment Discipline" pack can be mixed and matched. Agent A gets packs 1 and 3; Agent B gets packs 1 and 2. Each pack is version-controlled and editable without affecting the others.

Operations: Ingest/Query/Lint vs. Context/Work/Document

Karpathy defines three core operations for his wiki: Ingest (process new sources), Query (search and synthesize answers), and Lint (health-check the wiki for contradictions and gaps).

Waykee's workflow follows a different cycle: Get Context (receive full hierarchy), Do Work (write code, create tasks, send messages), and Update Documentation (reflect changes in Knowledge so future agents inherit the current state).

Philosophy: Personal Compiler vs. Team Context Engine

The deepest difference is philosophical. Karpathy's LLM Wiki is a personal knowledge compiler. It's designed for one person (or one agent) to accumulate and synthesize knowledge over time. The human curates sources; the LLM compiles and maintains the wiki. It's elegant, minimal, and self-contained.

Waykee Cortex is a team context engine. It's designed for organizations where multiple humans and multiple AI agents collaborate on shared work. Knowledge isn't compiled from raw sources — it's constructed through work and maintained as part of the workflow.

Where Karpathy Shines

• Personal research. Accumulating knowledge from papers, articles, and explorations over months.
• Solo projects. One developer maintaining context for their own work.
• Learning. Systematically building understanding of a new domain.
• Simplicity. No infrastructure needed — just markdown files and an LLM.
• Portability. The wiki is plain markdown. Works with any LLM, any editor, any platform.

Where Waykee Cortex Shines

• Team collaboration. Multiple humans and AI agents sharing context with granular permissions.
• Multi-agent orchestration. Different bots get different instruction packs and different levels of context.
• Integrated workflow. Knowledge, tasks, communication, and project management in one system.
• Construction history. See all tasks that have ever touched a component across all projects.
• Dual-parent context. A task inherits from both Knowledge (what it affects) and Work (when/why it's being done).
• External collaboration. Non-Waykee users participate via email, with messages integrated as chat bubbles.

Can You Use Both?

Absolutely. The two systems address different scopes and can complement each other:

• Karpathy's LLM Wiki for personal research. Use it to build personal expertise on a topic — ingesting papers, articles, and notes. The compiled wiki becomes your personal reference.
• Waykee Cortex for team execution. When research insights need to become organizational knowledge, create documentation in Waykee's hierarchy. Link tasks, assign agents, track progress.
• Bridge: export wiki → import as Waykee documentation. The compiled wiki pages can feed into Waykee's BotContext as documentation, making personal research available to the whole team through the hierarchy.

Technical Comparison

Conclusion

Karpathy's LLM Wiki and Waykee Cortex share a conviction: knowledge should compound, not be re-discovered. They diverge on scope. Karpathy built a beautiful pattern for a single mind. Waykee built an engine for entire teams.

If you're a solo developer or researcher accumulating expertise, Karpathy's pattern is hard to beat for simplicity and elegance. If you're running a team where multiple humans and AI agents need shared context, permissions, task tracking, and integrated communication — that's where Waykee Cortex earns its complexity.

The real insight from both systems is the same: stop treating AI as a search engine. Build persistent knowledge structures. Let the AI maintain them. Whether it's a compiled wiki or a hierarchical context engine, the future of AI knowledge management is compilation, not retrieval.