"The human's job is to curate sources, direct the analysis, ask good questions, and think about what it all means. The LLM's job is everything else."
— Andrej Karpathy

Based on Karpathy's LLM Wiki pattern, extended with:

• 🚫 Zero hallucination enforcement — every claim must cite a source
• 🤖 MCP server — expose your wiki to Claude Code, Codex, OpenClaw, Cursor
• 👥 Multi-agent support — multiple AI agents, one shared wiki, git handles conflicts
• 🔒 Human-in-the-loop conflict resolution — LLM flags contradictions, you decide truth
• 📊 Confidence tracking — per-claim certainty scores, quarantine mode for low-confidence pages

Why Not RAG?

Traditional RAG retrieves document chunks every query. Your LLM rediscovers knowledge from scratch. Every. Single. Time. Nothing compounds.

MeMex compiles knowledge once. Your LLM reads your sources and builds a structured, interlinked wiki. Every new source makes the whole wiki richer. Knowledge grows — it doesn't reset.

RAG:   Source → [retrieve] → LLM → Answer    (rediscovers every time)
MeMex: Source → [compile]  → Wiki → LLM       (compounds over time)

flowchart LR
    subgraph Input
        R["raw/ sources"]
    end
    
    subgraph Processing
        L["LLM reads & extracts"]
    end
    
    subgraph Knowledge
        W["wiki/ compiled knowledge"]
    end
    
    subgraph Output
        O["outputs/ reports"]
    end
    
    R --> L
    L --> W
    W --> O
    W -->|compounds over time| W

Key Principles

Quick Start

1. Clone

git clone https://github.com/JPeetz/MeMex-Zero-RAG.git my-wiki
cd my-wiki

2. Add a source

Drop any document into raw/:

cp ~/Downloads/interesting-article.md raw/

3. Tell your LLM to ingest it

Open Claude Code, Cursor, Codex, OpenClaw, or any AI coding agent and paste:

Read SCHEMA.md. Then ingest raw/interesting-article.md following the ingest workflow.

The LLM will:

• Read the source and discuss key takeaways with you
• Create a summary in wiki/sources/
• Create or update entity and concept pages
• Update wiki/index.md
• Log the operation in wiki/log.md

4. Query your knowledge base

Read wiki/index.md. Based on the knowledge base, answer: [YOUR QUESTION].
Cite which wiki pages informed your answer.

5. Run the MCP server (optional)

Expose your wiki to Claude Code, OpenClaw, or any MCP-compatible agent:

pip install mcp
python mcp/server.py --wiki wiki/

# SSE for remote/multi-agent access
pip install uvicorn starlette sse-starlette
python mcp/server.py --transport sse --port 3001

That's it. No setup scripts. No API keys. No database migrations.

Directory Structure

your-wiki/
├── L1/                      # 🔒 Auto-loaded, git-ignored (private)
│   ├── identity.md          # Your preferences, context
│   ├── rules.md             # Hard constraints, gotchas
│   └── credentials.md       # API keys (NEVER committed)
│
├── raw/                     # 📥 Source documents (immutable)
│
├── wiki/                    # 📚 LLM-maintained knowledge base
│   ├── index.md             # Master catalog
│   ├── log.md               # Operation history
│   ├── contradictions.md    # Pending conflict resolutions
│   ├── sources/             # One summary per ingested source
│   ├── entities/            # People, orgs, tools, projects
│   ├── concepts/            # Ideas, frameworks, patterns
│   └── synthesis/           # Analyses, comparisons, insights
│
├── outputs/                 # 📤 Generated reports, slides
├── mcp/                     # 🔌 MCP server (stdio + SSE)
├── scripts/                 # CLI tools (memex ingest/search/lint)
├── SCHEMA.md                # 🧠 The brain — how the wiki works
├── PROMPTS.md               # 📋 Copy-paste prompts for all operations
└── README.md                # This file

The L1/L2 Architecture

Inspired by CPU cache hierarchy:

L1 is git-ignored. It contains sensitive context that should never be committed.
L2 is the wiki. Versioned, shareable, grows over time.

Zero Hallucination Protocol

MeMex enforces citation discipline at every level:

1. Every claim must have a source: [Source: filename.md]
2. Lint catches violations: Unsourced claims are 🔴 ERROR, not warnings
3. Quarantine mode: Pages with >20% unsourced claims get status: quarantine
4. GitHub Actions: Broken links, missing citations, orphan pages checked on every PR

✅ Correct:  "Project X uses Redis [Source: architecture.md]"
❌ Rejected: "Project X uses Redis" (no source → lint error)

Conflict Resolution

When sources contradict each other:

1. LLM flags the conflict in wiki/contradictions.md
2. LLM stops and asks you which claim is authoritative
3. You decide — the LLM never auto-resolves truth
4. Decision is logged with rationale

MCP Tools

When running python mcp/server.py, your agent gets these tools:

Claude Code config (~/.claude/mcp.json)

{
  "mcpServers": {
    "memex": {
      "command": "python",
      "args": ["mcp/server.py", "--wiki", "wiki/"]
    }
  }
}

OpenClaw config

{
  "plugins": {
    "mcp": {
      "servers": [{ "name": "memex", "transport": "stdio",
        "command": "python3 mcp/server.py --wiki wiki/" }]
    }
  }
}

Multi-Agent Support

Multiple AI agents can write to the same wiki simultaneously using git worktrees:

# Each agent gets its own branch/worktree
git worktree add ../wiki-agent-1 feat/agent-1
git worktree add ../wiki-agent-2 feat/agent-2

# Agents commit with attribution
# wiki_write tool commits as: wiki(agent-name): add/update type/slug

Git handles authorship, history, and conflict detection natively. See GUIDE.md for the full multi-agent setup.

CLI Tooling

# Add to PATH
export PATH="$PATH:/path/to/memex/scripts"

memex ingest paper.pdf           # PDF with academic metadata
memex ingest https://example.com # Web page
memex search "machine learning"  # BM25 + semantic search
memex lint --fix                 # Check wiki health
memex graph                      # Generate wikilink graph
memex serve                      # Start MCP server

Requirements

• Any AI coding agent: Claude Code, Cursor, Codex, OpenClaw, Gemini CLI, etc.
• Git for version control
• Python 3.9+ for MCP server (optional)
• That's it — no databases, no APIs, no infrastructure

Optional Enhancements

Credits

• Vannevar Bush — Original Memex concept (1945)
• Andrej Karpathy — LLM Wiki pattern
• MehmetGoekce/llm-wiki — L1/L2 cache architecture inspiration
• rohitg00's LLM Wiki v2 — Lifecycle and consolidation patterns

Author

Joerg Peetz — with contributions from Molty, Coconut, Marvin (AI collaborators on the MeMex team)

License

MIT — Use it, fork it, adapt it, share it.

Start small. Ingest one source. Ask one question. Watch the wiki grow.

Part of the AgentForge Ecosystem

Copyright (c) 2026 Joerg Peetz. All rights reserved.