From LLM Wiki to AI-Native Memory
The LLM Wiki pattern points to a simple but powerful idea: AI should not just read your knowledge. It should help maintain, connect, and compound it.
Loading knowledge graph...
42 memories indexed.
3 new connections found.
Patterns: decisions, concepts, references
Wiki ready. Compounding active.
Background
What is the LLM Wiki pattern?
Andrej Karpathy introduced the concept of an "LLM Wiki" — a personal knowledge base where AI helps you capture, organize, and refine what you learn. The core insight: when you pair a structured wiki with LLMs, knowledge compounds. Each new piece of understanding builds on previous ones, and the AI helps you make connections you might miss.
The idea resonated with developers, researchers, and builders because it describes something many people already do in fragments: maintaining a personal knowledge base that grows richer with AI assistance. But the pattern also raised practical questions. How do you structure it? How do you maintain quality? How do you make it work across tools?
Key Insight
The Problem: RAG That Does Not Compound
Most RAG systems retrieve information, answer a question, and forget the result. That is useful, but it is not enough.
A real AI memory system should also preserve valuable outputs: decisions, comparisons, conclusions, explanations, and patterns discovered during work. That is the difference between retrieval and compounding knowledge.
Retrieval gives you back what you put in. Compounding turns every interaction into an opportunity to improve the knowledge base itself.
Comparison
LLM Wiki vs MEMM
| Capability | LLM Wiki Pattern | MEMM |
|---|---|---|
| Raw source preservation | Yes | Yes |
| AI-assisted synthesis | Yes | Yes |
| File-based knowledge | Yes | Yes |
| Local-first architecture | Conceptual | Product-level |
| AI tool integration | Manual | MCP/adapters |
| Context routing | Manual or custom | Built-in |
| Memory health/governance | Not defined | Built-in direction |
| Human-readable knowledge | Yes | Yes |
Implementation
How MEMM implements the LLM Wiki vision
Structured Memory Types
Knowledge, decisions, conventions, rules, preferences — each type gets its own structure and metadata, making retrieval precise and context-aware.
Context Levels (L0/L1/L2)
Not all knowledge is needed at all times. MEMM routes critical, relevant, and background context at the right granularity for each AI interaction.
Tool-Agnostic Adapters
Your LLM Wiki should not be trapped in one tool. MEMM translates structured memory into formats Claude, Cursor, ChatGPT, and Codex can consume directly.
Continuous Refinement
Every AI interaction is an opportunity to refine the knowledge base. MEMM is designed for knowledge that improves over time, not just retrieval.
FAQ
Common questions
What is an LLM Wiki?
An LLM Wiki is a personal knowledge base designed to be both human-readable and AI-usable. Popularized by Andrej Karpathy, the idea is that structured knowledge paired with LLMs creates compounding returns on what you learn and build.
How is MEMM different from a regular wiki?
A regular wiki is designed for humans to read and navigate. MEMM adds AI-native features: context routing, relevance scoring, tool adapters, and structured metadata.
Can I use MEMM as my personal LLM Wiki?
Yes. MEMM is a practical, local-first implementation of the LLM Wiki idea. You store knowledge in Markdown files, your AI tools access it through adapters.
Does MEMM replace my existing notes?
MEMM is not a general note-taking app. It is focused on knowledge that AI tools need as context. You can use it alongside Obsidian, Notion, or any other note system.
How does knowledge compound with MEMM?
When you use MEMM, every AI interaction can produce structured outputs (decisions, summaries, comparisons) that become new memories. Over time, the knowledge base gets richer.
Is MEMM local-first?
Yes. All memories are stored as local Markdown files. You own them, you version them, you move them. No cloud lock-in, no proprietary format.
MEMM turns the LLM Wiki idea into a working AI memory layer.
Open source. Local files. Compound your knowledge with every interaction.
Get MEMM