에이전트 강화
게시자가 제공하는 설치 방식입니다. PlugMCP 엔드포인트가 아닙니다.
원격 URL인가요, 내 컴퓨터에서 실행하는 명령인가요? 두 설치 방식의 차이
PlugMCP에서 실행
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
실행The Karpathy LLM Wiki pattern, production-ready.
No embeddings. No vector databases. No infrastructure. Just markdown, git, and your LLM.
Quick Start • MCP Server • Prompts • Schema • Examples
"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:
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
git clone https://github.com/JPeetz/MeMex-Zero-RAG.git my-wiki
cd my-wiki
Drop any document into raw/:
cp ~/Downloads/interesting-article.md raw/
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:
wiki/sources/wiki/index.mdwiki/log.mdRead wiki/index.md. Based on the knowledge base, answer: [YOUR QUESTION].
Cite which wiki pages informed your answer.
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.
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
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.
MeMex enforces citation discipline at every level:
[Source: filename.md]status: quarantine✅ Correct: "Project X uses Redis [Source: architecture.md]"
❌ Rejected: "Project X uses Redis" (no source → lint error)
When sources contradict each other:
wiki/contradictions.mdWhen running python mcp/server.py, your agent gets these tools:
wiki_flag / wiki_revalidate / wiki_read_by_slug)MeMex pages carry a status field (draft | active | stale | quarantine) so
agents and plugins (e.g. memex-hermes-plugin) can round-trip knowledge lifecycle
state without a translation layer. See SCHEMA.md
for the full field reference.
wiki_flag(page, reason) — quarantines a page (status: quarantine), stamping
the reason and a UTC timestamp into frontmatter. Refuses to flag a page written with
is_immutable: true.wiki_revalidate(page, outcome, new_body?, new_sources?, rationale?) — resolves a
quarantined page:
outcome: confirm — clears quarantine, sets status: active, keeps existing content.outcome: update — replaces the body with new_body (required) and optionally
appends new_sources citations, then sets status: active.outcome: retire — sets status: stale and clears flag metadata (reason/timestamp).wiki_read_by_slug(slug) — looks up a page by bare slug across all four type
directories (entities/, concepts/, sources/, synthesis/) so callers don't need
to know a page's type prefix ahead of time. Returns {"found": false} when missing,
or {"found": false, "ambiguous": true, "candidates": [...]} when the slug matches
more than one type directory.~/.claude/mcp.json){
"mcpServers": {
"memex": {
"command": "python",
"args": ["mcp/server.py", "--wiki", "wiki/"]
}
}
}
{
"plugins": {
"mcp": {
"servers": [{ "name": "memex", "transport": "stdio",
"command": "python3 mcp/server.py --wiki wiki/" }]
}
}
}
MeMex includes an optional knowledge graph that visualizes every [[wikilink]] in your wiki as an interactive, searchable force-directed graph.
# Enable when starting the MCP server
python3 mcp/server.py --enable-graph
The wiki_graph MCP tool builds graph/graph.html — open it in any browser.
The graph auto-rebuilds on every wiki_write when enabled.
Does not require --enable-graph to use the rest of MeMex — the graph is
an entirely self-contained optional feature.
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.
# 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
Joerg Peetz — with contributions from Molty, Coconut, Marvin (AI collaborators on the MeMex team)
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.
The Karpathy LLM Wiki, production-ready. Zero-RAG personal knowledge base with MCP server, multi-agent support, knowledge graph visualization, and hallucination enforcement. No embeddings, no vector D
개요
The Karpathy LLM Wiki, production-ready. Zero-RAG personal knowledge base with MCP server, multi-agent support, knowledge graph visualization, and hallucination enforcement. No embeddings, no vector DBs — just markdown + git.
도구
아직 도구 목록이 없습니다. 원격 서버는 첫 상태 확인 후에 도구 목록이 생깁니다.
FAQ
이 페이지에서 Cursor, Claude Desktop, Claude Code 또는 VS Code용 stdio 설정을 복사합니다. PlugMCP 프록시가 아니라 게시자의 설치 방식을 사용합니다.
이 레코드에는 필수 환경 변수가 등록되어 있지 않습니다.
MeMex Zero RAG은(는) stdio를 통해 로컬에서 실행됩니다.
마지막 확인: 2026-10-28. The Karpathy LLM Wiki, production-ready. Zero-RAG personal knowledge base with MCP server, multi-agent support, knowledge graph visualization, and hallucination enforcement. No embeddings, no vector DBs — just markdown + git.