Source: ai-research/agentwikis-platform-overview-2026-06-12.md — scraped from agentwikis.com 2026-06-12

Agent Wikis (agentwikis.com) is a commercial service that builds and maintains structured knowledge bases using the Karpathy LLM-wiki pattern, then publishes them as free, agent-readable Markdown with llms.txt indexes. Built by a developer on the Hermes Agent team — the Hermes wiki is the service’s flagship and the research backbone for their YouTube video series. Agent Wikis is the first commercially-operated multi-wiki Karpathy implementation with published empirical accuracy benchmarks proving why compiled wikis outperform raw-source RAG.

Key Takeaways

  • Compiled wiki vs raw-source RAG — same model, same retrieval, same token budget: compiled wiki hits 89% correct / 7% hallucination vs raw-source RAG’s 63% / 26%. Maintenance (curation) is the lever, not retrieval method.
  • vs. web search — web search scores 48% / 48% on niche fast-moving domains; compiled wiki is reproducible where search varies ±10–14 points between runs. On synthesis/decision questions: web scored 0%, wiki held 100%.
  • Karpathy attribution — explicitly credits Karpathy’s GitHub gist and llm-wiki-template; README on each wiki is the standard Karpathy pattern (raw/ → wiki/ → CLAUDE.md schema).
  • Agent-native architecture — every wiki exposes llms.txt, llms-full.txt, index.json, and an MCP server (calibrated search + section-level reads) — the exact configuration that scored 89% in their eval.
  • Multi-domain catalog — 15 live wikis across AI agents, inference/serving, image/video gen, DeFi, trading, and paid ads (2026-06-12).
  • Routing beats either source — “wiki first, web on gaps” scored 93%, above either source alone.

The Benchmark in Full

27 tasks on Hermes Agent (releasing every 3–5 days), same model and token budget, blind LLM judge, replicated across two model families:

ConditionCorrectHallucination
Parametric (no context)4%85%
Web search48%48%
RAG over raw sources63%26%
RAG over compiled wiki89%7%
Whole curated pages (static)81%22%
Agent navigates wiki70%30%

Cost: ~0.0054/query (~3x cheaper than web search), ~1.2s vs ~3.6s (~3x faster).

Wiki Catalog (2026-06-12)

AI Agents: Hermes (105 docs), Claude Code (33), Codex (36), Agent Workflows (15) Inference/Serving: llama.cpp (63), vLLM (19), LM Studio (23) Image/Video Gen: ComfyUI (31), HyperFrames (167), Remotion (31) DeFi: Hyperliquid (18), Solana (16) Trading: Trading (44), Stock Options (16) Ads: Google Ads (27), Meta Ads (25)

How It Differs from Other Karpathy-Pattern Implementations

ImplementationScaleAccessBenchmarks
Stride starterSingle vault, personalObsidian onlyNone
SynthadocSingle vault, self-hostedObsidian + APINone
Matt Wolfe second brainSingle vault + CRM/journalCodexNone
Agent Wikis15+ wikis, hosted serviceWeb + llms.txt + MCPPublished (89% vs 48%)

The key distinction: Agent Wikis is the only implementation that (1) runs as a service maintaining many wikis simultaneously, (2) exposes an agent-native access path without requiring users to run their own LLM loop, and (3) has published a controlled eval proving the pattern’s accuracy advantage over web search.

Pricing

  • Free — browse all public wikis, raw Markdown + llms.txt, no account required
  • Sponsor — $50 one-time, you pick the topic, published as free public wiki
  • Private20/mo, compiled from your private sources, delivered as private git repo
  • Project100/mo, claim the public wiki for your product, includes eval report

Try It