RAGLiving Wiki

Karpathy (April 3, 2026): "RAG rediscovers everything from scratch on every query." His alternative: a Living Wiki where AI accumulates knowledge over time. We transitioned from RAG to Wiki in production.
Based on Karpathy's update — covered by VentureBeat, TechBuddies, DAIR.AI

Our Transition

Jan 2026
Arena Brain v1
Classic RAG
Dump all data
Mar 2026
Living Wiki created
wiki_compiler.py
Lessons accumulate
3 Apr 2026
Karpathy publishes
"RAG is outdated"
Wiki pattern
5 Apr 2026
Arena Brain v2
Graph-augmented
Wiki-first retrieval
6 Apr 2026
Meta-Harness
Decay + consolidation
Wiki evolves itself

❌ Classic RAG

Every query starts from zero

0
queries
0
learned
VS

✅ Living Wiki

Knowledge accumulates over time

0
queries
0
accumulated

Live Wiki Data

Our Living Wiki currently contains:

Loading wiki stats...

Beyond Karpathy: LLM Wiki v2.0 — The Strategic Layer

Karpathy's wiki accumulates knowledge. Our extension generates testable hypotheses. The wiki doesn't just record what worked — it thinks about what to try next and sends instructions to the Darwin Engine.

Three layers beyond Karpathy's v1.0:

1. RAW
8 engines generate raw JSON every night
2. COMPILED
Wiki Compiler structures into Markdown
3. STRATEGIC (v2.0)
Generates hypotheses → Darwin tests them
See the Wiki Grow Live Read the Full Article Knowledge Graph

Strategy Arena transitioned from classic RAG to a Living Wiki pattern following Andrej Karpathy's April 2026 insight. Classic RAG (Retrieval Augmented Generation) searches a vector database on every query, rediscovering information from scratch each time. The Living Wiki pattern compiles raw data into structured Markdown documents that accumulate intelligence over time. Strategy Arena's implementation adds graph-augmented retrieval (inspired by LightRAG), decay + consolidation (inspired by elfmem), and a Meta-Harness agent that optimizes the wiki compiler itself. The result: 2,500+ experiments compiled into actionable lessons, connected through a Knowledge Graph of 121 AI nodes. This is not theoretical — it runs in production every night on 59 live AI trading strategies.

Rejoindre le canal 💬 Feedback ☕ Support Us