Karpathy's LLM Wiki Is a Manifesto for What We Already Built — And the One Gap We Just Closed

[xg-gh-25]

Karpathy's LLM Wiki Is a Manifesto for What We Already Built

"The wiki is a persistent, compounding artifact. The cross-references are already there. The contradictions have already been flagged. The synthesis already reflects everything you've read."
— Andrej Karpathy, LLM Wiki

Karpathy published his LLM Wiki pattern this week. It describes exactly what SwarmAI has been running in production for 3 months — and names the one gap we hadn't closed yet.

---

The Pattern (30-second version)

Instead of RAG (stateless retrieval on every query), the LLM builds and maintains a persistent wiki — structured, interlinked markdown. Knowledge compiled once, kept current. Three layers:

Layer	What	Who owns it
Raw sources	Immutable inputs	Human curates
Wiki	Structured, cross-referenced pages	LLM writes & maintains
Schema	Rules for how the wiki works	Human + LLM co-evolve

Three operations: Ingest (source → wiki pages), Query (ask → synthesize → answer), Lint (health-check contradictions, orphans, gaps).

---

How SwarmAI Already Implements This

Karpathy's Concept	SwarmAI Equivalent	Status
Raw sources	Knowledge/ (72 DailyActivity + Notes + Learned + Reports)	✅ Production 3 months
Wiki (LLM-maintained)	MEMORY.md + EVOLUTION.md + DDD docs (4 docs × 8 projects)	✅ Production
Schema	11 context files (AGENT/SOUL/STEERING/TOOLS/...)	✅ Production
Ingest	Post-session hooks + DDD Cultivation Engine (event-driven, 6 sources)	✅ Production
Query	Smart memory injection + FTS5 recall + session resume enrichment	✅ Production
Lint	s_loops-health (31 checks) + context_health_hook	✅ Production
Code graph	Code Intelligence (AST parser + graph_store + REST API)	✅ Production
Evolution	L0-L3 governance + Evolution Pipeline v2	✅ Production

We go further than the manifesto in several dimensions:

• Self-evolution — the wiki doesn't just accumulate, it upgrades its own rules (SOUL.md, AGENT.md)
• Autonomous pipeline — 8-stage TDD + adversarial review driven by wiki knowledge (DDD)
• Proactive intelligence — push-based briefing, not just pull-based query
• Multi-project isolation — each project has its own 4-doc knowledge substrate
• Mechanical gates — hooks enforce wiki maintenance (not honor-system discipline)

---

The One Gap: Query Result Backflow

Karpathy's key insight we hadn't implemented:

"Good answers can be filed back into the wiki as new pages. A comparison you asked for, an analysis, a connection you discovered — these are valuable and shouldn't disappear into chat history."

We had everything EXCEPT this. Session insights lived in chat history and DailyActivity logs but never flowed back as first-class Knowledge pages.

Shipped today: KnowledgeBackflowHook — a post-session hook that:

1. Scans all assistant messages for high-value outputs (>500 prose words + structural analysis markers)
2. Auto-captures qualifying outputs as Knowledge/Notes/YYYY-MM-DD-<slug>-<hash>.md
3. Atomic writes, YAML-safe frontmatter, content-hash dedup, max 3 per session
4. Fires between DailyActivity extraction and auto-commit (gets committed automatically)

---

Bonus: Patterns Borrowed from Graphify + ΩmegaWiki

We also studied two implementations that took Karpathy's idea in different directions:

Project	Stars	What it does	What we borrowed
Graphify	55.7K	Code → knowledge graph (18 platforms)	P3: Git hook auto-rebuild — incremental re-index on commit
ΩmegaWiki	874	Research lifecycle platform (24 skills)	P2: Anti-repetition memory — failed experiments block similar approaches

P2: Anti-Repetition Check (from ΩmegaWiki)

ΩmegaWiki treats failed experiments as "anti-repetition memory" — not just archived, but actively blocking similar future directions.

We added this to our Pipeline EVALUATE stage: before recommending GO, the stage now explicitly cross-references IMPROVEMENT.md "What Failed" for structurally similar approaches. Must cite the failed entry and explain why this attempt is different — or REJECT.

P3: Code Intel Auto-Rebuild (from Graphify)

Graphify uses git hooks to rebuild the knowledge graph after every commit (AST-only, zero API cost). Fresh graph with zero human effort.

We extended our existing CodeChangeFeed post-session hook to also re-index changed .py/.ts files into code_intel.db. The graph stays fresh automatically.

---

What We Explicitly Did NOT Borrow

Feature	From	Why we skipped
Wiki-from-graph generation	Graphify --wiki	Our TECH.md (104K) already IS the architecture wiki. Two systems = drift.
9 entity types	ΩmegaWiki	Over-classification. DDD 4-doc is more flexible.
Neo4j export	Graphify	SQLite graph sufficient at our scale.
Multi-platform support (18 IDEs)	Graphify	We're one integrated system, not a tool library.
LaTeX paper pipeline	ΩmegaWiki	We don't write papers.

---

The Fundamental Difference

	Karpathy's Manifesto	Graphify	ΩmegaWiki	SwarmAI
Knowledge form	Flat markdown wiki	JSON graph + viz	9-type entities + edges	DDD 4-doc × N projects
Automation	None (manual)	Git hook rebuild	GitHub Actions cron	Daemon + jobs + hooks (24/7)
Self-evolution	❌	❌	❌	✅ L0-L3 governance
Quality gates	Lint (passive)	None	Review LLM	Pipeline + adversarial + convergence

SwarmAI is the only system where the wiki improves its own rules. Others are stationary. We are non-stationary.

---

Related Discussions

• #19 — AI Agent 不需要 Neo4j: 知识管理的达尔文主义 — the ontology philosophy behind our knowledge lifecycle
• #51 — AI-Ready Repo Standard (EN) — the .ai-context/ convention for making repos AI-consumable
• #52 — AI-Ready Repo Standard (CN) — 中文版
• #31 — The Personality Trap — why governance > personality for agent compliance
• #32 — Multi-Specialist Adversarial Review — the review system that caught 3 HIGH bugs in today's "trivial" hook

---

Source References

• Karpathy's LLM Wiki gist: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
• Graphify (55.7K ⭐): https://github.com/safishamsi/graphify
• ΩmegaWiki / AutoSci (874 ⭐): https://github.com/skyllwt/AutoSci

---

The manifesto describes first principles. The tools implement subsets. SwarmAI runs the full loop — and the loop improves itself.