The Graph
Nodes and edges. Binary connections. This is what Obsidian, Neo4j, and the entire RAG industry runs on. Useful but shallow. Cannot encode why two things connect or how strongly.
EST. MMXXVI · VOL. I · A LIVING WORLD MODEL
The Giga Library.
The Giga Library.
Every fact knows where it came from,
when it was true,
and who disagreed.
WikiDesignCo is the reasoning substrate beneath your agents. A temporal, multi-perspectival, epistemically rigorous metagraph stitched together from thousands of textbooks, hundreds of channels, and the entire shifting frontier of research. Built for the operators who are done pretending RAG is enough.
§ I · The Archive
A library that reads itself
Imagine the Library of Alexandria if every scroll had embeddings, every claim had a temporal validity window, and the librarians were AI agents that argued with each other in the margins.
Obsidian stores notes. Notion stores docs. Pinecone stores vectors. WikiDesignCo stores understanding. Every entity carries Dwarf Fortress-level detail: provenance, relationship history, confidence scores, temporal evolution, and cross-domain resonance.
VOL. IKnowledge Ingestion
VOL. IIEpistemic Stratification
VOL. IIIBi-Temporal Validity
VOL. IVThe Contradiction Engine
VOL. VHybrid Retrieval
VOL. VIAgent Redwood
VOL. VIITokenomics
VOL. VIIIContent Studio
VOL. IXMeme Shaman
VOL. XWorkflow Trace
§ II · The Substrate
Graph → Hypergraph → Metagraph
Most "knowledge graphs" are filing cabinets with extra steps.
A graph says A is connected to B. A hypergraph says {A, B, C, D} share this property. A metagraph says the relationship between {A, B, C, D} is itself a fact that other relationships can point to.
The Hypergraph
A single edge can connect any number of nodes. An entire research consensus, a multi-author paper, a panel discussion, all become first-class objects. The relationships start having shape.
The Metagraph
Edges point at edges. "Source S₂ revises the claim that source S₁ made about concept C" is a single fact. Every claim arrives with provenance, confidence, and an audit trail. Reasoning becomes structural.
§ III · The Accession Desk
Ingestion as industrial process
Books arrive. They are read by twelve agents. They leave as living knowledge.
live conveyor
Four streams, one substrate.
Textbooks tell you what is established. YouTube tells you what works in practice. Research papers tell you what is becoming possible. Social tells you what the frontier is reacting to. Each gets nested fan-out/fan-in processing: chapter agents → sub-agents (entities, relations, resonance) → fan back into a book-level graph that is then cross-referenced against the existing metagraph.
- spaCy + custom NER · Rebel relation extraction · coreference resolution
- Three-layer entity resolution: exact → embedding → LLM disambiguation
- Every claim writes a provenance chain: source → claim → entity
- Graphiti episodes auto-close validity windows on superseded claims
accession pipeline · six stages
diagram
books
youtube
papers
social
extractok
nlpok
resolveok
graphok
contradictok
embedok
streams: books · youtube · papers · social
stages: extract → nlp → resolve → graph → contradict → embed
§ IV · The Triangle of Grounding
Three pillars of knowledge
The best answer often lives in the gap between what the textbook says, what practitioners actually do, and what the paper just discovered.
Textbooks
What the field has agreed upon. Peer review, structured pedagogy, the slowest but most reliable signal. The foundation layer of the library.
YouTube
Practitioner knowledge. Every video transcript gets multi-agent semantic analysis. The gap between textbook and practice is where the gold is.
Research
The bleeding edge. Where the field is going before textbooks update. We parse the citation graph automatically, so every paper lands as a node in a living argument structure.
§ V · The Date-Stamped Stack
Bi-temporal validity
Every fact knows when it was true.
Standard RAG retrieves the highest-similarity chunk, even if the author has since contradicted themselves three times. Graphiti gives every fact two timestamps: t_event (when the world said it) and t_ingestion (when we learned it). Scrub the timeline and the world model re-asserts.
temporal scrubber
Time is a first-class column.
An author's 2019 position contradicts their 2024 one. Standard graphs overwrite. WikiDesignCo closes the valid_to on the 2019 claim, opens a new one at 2024, and keeps the entire history queryable.
- Point-in-time queries: "what would I have known on Jun 1 2023?"
- Diff: surface every position-shift across an author's corpus
- Validity windows propagate through dependent claims automatically
temporal.scrub · karpathy / LLM scaling
live
20172019202120232024
§ VI · The Marginalia
The contradiction engine
When two sources disagree, the system surfaces the controversy and leaves it standing.
contradiction detected
The disagreement survives retrieval.
Every new claim runs the conflict-detection query. Overlapping entity sets + opposite predicates + overlapping validity windows = a Controversy node. Agents querying that topic receive the controversy structure, not a false-confidence average. They can present it, resolve it, or hedge, and they always know.
- Classifies: factual · definitional · contextual · temporal
- Routes definitional disputes to ontology clarification
- Contextual disputes get applies_when conditions auto-attached
- PLN-style truth values: (strength, confidence) two-vector epistemic state
contradiction.live
unresolved
CLAIM A · 2017
"Transformers categorically outperform RNNs."
CONTROVERSY · #C-0421
"Transformer dominance"
5 supporting
↔ contextual dispute
3 contradicting
CLAIM B · 2024
"Mamba / SSMs match transformers at long-sequence tasks at <1/3 compute."
§ VII · The Reading Room
The Giga Librarian
Not keyword search. Not vector similarity. An agentic librarian that understands what kind of question you're asking.
Type a query. The Giga Librarian decides whether you need established theory, practical implementation, or emerging research, then synthesizes across all three through the same retrieval, weighted by epistemic type.
hybrid retrieval
Vector, graph, and BM25 in one SurrealQL pass.
The default query hits all three retrieval modes simultaneously: dense vector for semantic, graph traversal for relationship-based context, and BM25 for keyword precision. Merged, re-ranked, returned with full provenance.
- Cosine sim against task-specific Google Embeddings 2 modes
- Graph-walked sub-graph centered on query entities, 2-3 hops out
- BM25 hybrid over SurrealDB / Elasticsearch full-text index
- Result confidence weighted by source-type credibility chain
retrieval.lab
live
Vector · semantic
GeLU outperforms ReLU in transformer FFN
SwiGLU activation gains in PaLM
Activation choice and gradient pathology
Graph · 2-hop
Karpathy: SwiGLU in production at OpenAI
Shazeer: GLU variants ablation
Llama 3 architecture notes
BM25 · keyword
"hidden layer activation transformer"
"GeLU GELU activation function"
"FFN feed forward network"
§ VIII · The Architecture
Four layers, one substrate
A stack chosen for what stays up. Pragmatist by construction.
We use OpenCog Hyperon and AtomSpace as design reference, then implement metagraph semantics in production-grade tools we can actually operate: Graphiti on Neo4j with SurrealDB unifying vector + graph + full-text. PLN-inspired confidence propagation in our own reasoning layer.
04
agent interface
Agent Redwood · 12 components
Every agent gets three world-model tools: query(), assert(), update_belief(). Episodic memory, planning, constraints, reward, and strategy all live as subgraphs in the same metagraph, so reasoning across them runs on the structure itself.
03
reasoning + retrieval
Graphiti · pgvector · BM25 hybrid · PLN-lite
Bi-temporal knowledge engine. Hybrid retrieval lab fuses vector, graph traversal, and full-text in one query. Custom confidence propagation lets truth values flow through dependent claims when new evidence lands.
02
knowledge graph core
Neo4j + APOC · SurrealDB unified
Metagraph patterns implemented as graph patterns: epistemic nodes (reified claims), truth-value edges, provenance chains, controversy nodes, belief networks. SurrealDB collapses the duct-tape stack: graph, vector, doc, full-text all in one engine, one query language.
01
ingestion pipeline
Scrapers · NLP · ontology alignment
Real-time scrapers (Twitter, RSS, YT) + batch ingestion (PDF textbooks, papers). NER, relation extraction, coreference, entity dedup, conflict detection. Every claim enters with its provenance attached to the structure itself.
§ IX · The Reading Apprentices
Agent Redwood · 12 components
Twelve layers. One shared substrate.
In most agent frameworks the twelve layers are structurally isolated. Memory doesn't inform reward, strategy doesn't query constraints. In Redwood they all live as subgraphs in the same metagraph. The agent reasons across them without explicit plumbing.
- personality · planning · mission · constraints
- memory · evaluation · tools · awareness
- reward model · metadata · strategy · integrations
- each is a subgraph; cross-layer inference is automatic
An agent with "avoid asserting contested claims" can look up which topics have low epistemic confidence and adjust language, without the constraint module and the knowledge module being explicitly wired. They share the world model. The inference is structural.
Agent coreshared world model
§ X · The Competitive Stacks
Why this beats RAG, KMs, and vector DBs
RAG is a bicycle. This is a car.
Bicycles have their place. You don't race one on a highway.
| Multimodal embeddings | NLP enrichment | Edges-of-edges | Bi-temporal | Contradiction | Agent-native | Token economy | |
|---|---|---|---|---|---|---|---|
| WikiDesignCo | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Notion | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Obsidian | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Roam / Logseq | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Pinecone / Weaviate | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Microsoft Copilot + Graph | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Standard RAG stack | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
§ XI · The Lending Rules
Credits · subscriptions · stakes
Burn-and-mint, not seat-and-bill.
Queries burn credits. Verified enrichments mint them back. Subscribers get the best per-credit price, plus a stake in the graph their work helped build.
SCHOLAR
Individual
$15/mo
For researchers, writers, and knowledge workers who outgrew Obsidian.
- 10,000 entities · personal graph
- Basic NLP enrichment pipeline
- 2,000 query credits / month
- Notion · Linear · Drive connectors
CONSTELLATION
Team
$40/seat/mo
Shared graph. Agent integrations. Workflow tracing. The full Giga Library experience.
- 100,000 entities · shared namespace
- Full NLP + agent pipeline
- 20,000 query credits / seat / month
- Workflow run explorer + review inbox
- Studio: storyboard, voice, motion
SOVEREIGN
Enterprise
Custom
For organizations running multi-agent operations across business lines.
- Unlimited entities · isolated VPC
- Custom enrichment pipelines
- SSO · SLA · DPA · audit
- Token-economy onboarding
§ XII · The Reference Desk
Common questions
The questions people actually ask.
Is this just another wrapper on top of vector search?
No. Vector retrieval is one of three modes in the Giga Librarian. The substrate beneath is a metagraph: edges point at edges, every claim carries provenance and a bi-temporal validity window, and a continuous contradiction engine surfaces controversies instead of averaging them away. We use embeddings; we are not made of embeddings.
Why Graphiti / Neo4j and not OpenCog AtomSpace?
AtomSpace is research infrastructure, correct and expressive, and missing the operational amenities you need to ship. We treat AtomSpace and Hyperon as a design reference and implement the same principles (typed atoms, truth values, edge-to-edge relations) in production-grade tooling we can actually operate. Strategy B in the architecture doc.
How does the token economy work without being a security?
Credits operate identically to SaaS at the surface: pay for a subscription, get credits, burn them on queries. Contributors who add enrichments that pass quality gates mint credits back. The Solana layer is additive: if regulation forecloses tokenization the SaaS layer operates independently. We do not promise yield; credits are utility.
Can I bring my own corpus?
Yes. The ingestion pipeline accepts PDFs, EPUBs, YouTube channels, RSS feeds, Twitter/X handles, and direct Markdown/Notion exports. We ontology-align against your existing canonical entities or seed a fresh namespace. Bronze → Silver → Gold → Platinum → Diamond progression as quality compounds.
What happens when two sources disagree?
A Controversy node is created automatically, classified (factual / definitional / contextual / temporal), and routed to the appropriate resolution strategy. Agents querying the topic receive the controversy structure, supporting count, contradicting count, current resolution status, not a hallucinated synthesis.
Is there an MCP server?
Yes. The Giga Librarian exposes query / assert / update_belief over MCP, plus subscription endpoints for episode streams. Agent Redwood instances connect natively. Custom agents via Claude Code, Cursor, or your own framework, same contract.
A LIVING WORLD MODEL · BUILD WITH US
Step inside the stacks.
Almost no one is building metagraph-based world models for production agentic systems right now. The tools exist. The theory is established. The accumulated episodic data is the real moat, and it compounds from the day you start writing.