Neural Fabric · Knowledge Graph

How governed business knowledge reaches every agent.

Add business data once. Qorbit structures it into entities, derives the relationships, governs every write, and exposes the same graph through MCP.

Input
Documents, entities, and agent-written decisions.
Fabric
Governed graph construction with provenance.
Agents
Same eight MCP tools in every connected client.

01

Write once. Recall anywhere.

Write from any MCP-connected agent. Ask from any other. The record returns with date, confidence, and source.

Agent A · IDE
// Write a governed business decision fabric_add({ name: "Teller vs Plaid decision", type: "Decision", properties: { summary: "Chose Teller over Plaid — simpler API, better webhooks.", date: "2026-02-14" } }) // → ✓ Persisted to Neural Fabric · governed write
14 days later · different agent
// Recall from any connected agent fabric_search({ query: "Why did we pick Teller?" }) // → Decision: Chose Teller over Plaid — simpler API, better webhook reliability. // source: Decision · Feb 14, 2026
 
Neural Fabric
LLM Agent
-
✓ saved to Neural Fabric
AI Agent
-
✓ saved to Neural Fabric
Document
-
✓ saved to Neural Fabric
Any Agent · Later
"What do we know about this project?"
Synthesised answer · 3 sources
confidence 1.0 · 812ms
LLM
Agent
Client wants the dashboard delivered by end of March.
written 6 days ago
AI
Agent
Budget approved. Team aligned on delivery scope.
written 3 days ago
Document
Spec doc, 5 requirements connected to this project
ingested 8 days ago
↗ graph traversal · not text search · 3 sources · grounded in the Neural Fabric

02

What makes it different.

Zero embedding cost
At ingestion
No external embedding provider. No per-ingestion API call to encode your data. Frequency-native encoding runs 250× faster than dense embeddings.
Multi-hop
Graph traversal
Answers emerge from connections across your entire knowledge graph, not from a single matching chunk. Follows chains of related entities standard retrieval can't see.
Governed refusal
Anti-hallucination
When the evidence isn't there, the Neural Fabric returns "No information available", not a guess. Gets quieter, not wronger. Every answer traces to its source.
Bring your own LLM
Your keys, your provider
Document extraction runs on your Anthropic, OpenAI, or compatible API key. Your API key is encrypted at rest with AES-256-GCM. Used only for extraction and classification, never exposed through any API.

03

Eight tools. Every agent.

Connect any MCP-compatible AI tool and access your entire knowledge graph through eight governed tools. No SDK. No integration code.

fabric_add
Add entities and documents to the graph
fabric_edit
Partial, targeted modifications to existing entities
fabric_search
Search by name, type, or natural language
fabric_list
List and filter entities across your graph
fabric_graph
Entity neighborhoods and resonance subgraphs
fabric_read_source
Read source document content from stored files
fabric_delete
Remove entities with tombstone semantics
fabric_provenance
Cryptographic audit trail with tamper detection

Same eight tools in Claude, ChatGPT, Cursor, Claude Code, or any MCP-compatible agent.

04

Validated at scale.

Embedding-based retrieval degrades as data grows. Vector spaces lose discriminative power as they get crowded: more entities mean more noise, more false positives, more re-ranking layers to compensate. The Neural Fabric doesn't have this failure mode. Relationships are mathematically derived, not statistically approximated. The spectral radius is constrained below 1.0 at every write.

Validated at 60,000+ entities

Validated on reproducible synthetic corpora with zero steady-state errors.
123 sustained mixed-workload operations at 60K entities, zero errors.

1.0
identity recall
Across 124,000+ verified pairs.
If a connection exists in your documents, your agent finds it.
1.0
Recall@10
Target identity appears in the top 10 every time.
Ask for a customer by name. The right record is in the top results every time.
0.925
MRR
Correct identity ranks near the top.
The correct entity is usually first. Less hunting, less wrong context.
0.993
Precision
False positives stay constrained.
Under 1% false connections. Every edge shows its evidence.
$0
embedding cost
No vector API calls.
Ingestion costs you nothing in tokens. Competitors meter ingestion because their construction runs on LLMs.
None
re-ranker
No compensation layer.
5,188
ENTITIES/SEC
2,149 under default config, 5,188 with aggressive preset. Frequency-native encoding, not dense vector computation.
120B
per entity
In-memory propagation footprint. Linear scaling validated to 60,000+ entities.
1,097×
vs dense at 10K
Sparse CSR vs dense matrix propagation time at 10K entities. Dense: 47.5s. Sparse: 43.3ms.
0.998
SPECTRAL RADIUS
Constrained below 1.0 at every write. Every write is validated for mathematical stability before it lands.
99.8% candidate reduction at every scale tested. Multi-hop traversal verified at 10 hops. Zero embedding cost at any scale.
Your agent reads a shortlist, not a corpus. Smaller contexts, fewer tokens per answer.

Governed knowledge graph for multi-agent business knowledge.