STRUCTURED ENGINE

mindBrain
The structured engine beneath GhostCrab.

mindBrain is a structured agentic database. It can store any content — the underlying engine is PostgreSQL (or SQLite in the Personal edition), so any table structure is possible: contacts, contracts, tickets, tasks,
documents, whatever the domain requires.

What makes mindBrain different is the meta-layer that sits above that content.

INSTALL ON GITHUB READ THE ARCHITECTURE SEE USE CASES
GhostCrab mascot

Hello, I'm GhostCrab — your favorite MCP server.

WHAT mindBrain IS

Coherent world model for agents.

Every dataset — regardless of its structure — is connected to three capabilities: Facettes for multi-criteria retrieval, Graphes for typed directed relationships, and Projections for compact working context. That meta-layer is always there. It is what makes the content navigable for an agent, not just queryable by a developer.

Official site: mindbrain.be.

Without mindBrain, an agent operates on raw data or reconstructed prose. With mindBrain, it operates on a
coherent, traceable, controlled world model — one it can query, traverse, and reason over in bounded time.

TWO VERSIONS

Two versions. One architecture.

mindBrain is available in two versions depending on the scale of the project.

mindBrain Personal

Engine
SQLite
For
Individual use, small teams, personal projects
Setup
Single file, zero infrastructure
Performance
Fast for thousands of entities
Throughput
Lightweight, embedded
Good for
CRM personnel, family projects, research, side projects

mindBrain Professional

Engine
PostgreSQL
For
Teams, enterprise, high-volume workloads
Setup
Self-hosted or managed, production-grade
Performance
Millisecond filtering on tens of millions of documents
Throughput
~4.3B documents per table (Roaring Bitmaps on 32-bit IDs)
Good for
Onboarding at scale, compliance, multi-team coordination

Both versions expose the same three capabilities. The architecture is identical. Only the engine changes.

THREE CAPABILITIES

mindBrain exposes three distinct capabilities. GhostCrab makes all three accessible to any MCP-compatible agent.

01 Facets

How the agent finds the right slice of a domain.

Instead of reading everything, the agent narrows the field — by status, owner, country, phase, priority,
or any dimension that is meaningful to the project. Facettes uses precomputed bitmap indexes and hybrid
BM25 + embedding retrieval to return the relevant subset in milliseconds.

The agent does not scan. It queries a deterministic interface and receives a bounded result.

Example questions

  • → Which onboarding cases are blocked in Belgium?
  • → Which deals are in negotiation with no legal review yet?
02 Graphs

How the agent understands dependencies, blockers, and missing knowledge.

REQUIRESBLOCKSVALIDATESDEPENDS_ON

Direction matters. A → B is not the same as B → A. Professional domains are full of asymmetric relationships:
legal prerequisites, onboarding workflows, diagnostic protocols where order is non-negotiable.

Graphes also detects structural gaps: by comparing the current graph against the domain ontology, it infers
that edges should exist. What was invisible becomes an explicit signal.

03 Projections

How the agent gets a compact working context for the current task.

FACT GOAL STEP CONSTRAINT

Even a perfectly modeled project cannot be dumped wholesale into an agent's context window. Projections
pre-pack the right FACT, GOAL, STEP, and CONSTRAINT for the current task — ranked, formatted, with provenance attached.

Typically 80 to 200 bytes. No cold-start search. No transcript replay. A ready-to-use reasoning surface for this turn.

WHY THIS ARCHITECTURE EXISTS

Most agent failures follow the same pattern.

  • ✗ too much raw context
  • ✗ no explicit structure for known vs. unknown
  • ✗ no dependency model
  • ✗ no separation between source data and working context
  • ✗ no durable project state across sessions

Each of the three mindBrain capabilities addresses a distinct failure mode. Facettes solves retrieval.
Graphes solves structure and dependency. Projections solves context packaging.

Together they form a coherent stack — not a collection of memory plugins, but a purposeful architecture
for agents that work on real projects over time.

GHOSTCRAB IS THE INTERFACE

mindBrain is the engine. GhostCrab is the door.

GhostCrab is an open-source MCP server that exposes all three mindBrain capabilities — Facettes, Graphes,
Projections — as tools any MCP-compatible agent can call. Claude Code, Cursor, Codex, or any agent that speaks
MCP can use GhostCrab without changing its interface.

The agent still does the reasoning. GhostCrab still handles conversation and intent. mindBrain holds the structure. Each layer does exactly one thing.

Your agent → reasons, interprets, proposes ↓ GhostCrab (MCP) → exposes structure as callable tools ↓ mindBrain → holds entities, graphs, projections, state ↓ SQLite / PostgreSQL → durable storage

The agent never touches the raw database. It calls GhostCrab tools. GhostCrab queries mindBrain.
mindBrain returns a bounded, typed, structured answer. No hallucinated queries. No unbounded scans. No cold-start reconstruction.

05 / Ubiquity

The graph structures reality

Wherever there are objects and relationships between them, there is a graph. CRM, legislation, project management, mental models — the structure is always the same: nodes, edges, weights, facets.

// CRM

Account → Contacts → Opportunities → Activities. Every entity is a node. Every relationship is a typed edge.

// Law / regulation

Article → Paragraph → Cross-reference → Case law. Legal text is a hypergraph of references.

// Project management

Tasks → Dependencies → Resources → Deliverables. The critical path is a weighted graph problem.

// Mental model

Concepts → Causal links → Analogies → Exceptions. An expert's ontology is a knowledge graph.

06 / The problem

A raw graph is blind at scale

Graph structure answers "how objects are connected". It does not answer "which objects match these criteria at once". That is what is missing.

⚡ The core issue

To answer a question like "all opportunities >€50k for tech-sector accounts with an active contact in Brussels", a standard graph must traverse every relevant node and edge. On a real graph, that is not a query — it is an exploration.

No index Naive traversal
O(V + E)
  • Visits every potentially relevant node
  • Evaluates every edge one by one
  • 1M nodes = 1M operations minimum
  • The agent waits, retries, hallucinates on timeout
  • Hard to parallelize without conflict
pg_facets Precomputed indexes
O(1) lookup
  • Precomputed Roaring bitmaps per facet
  • Set intersection = AND across criteria
  • Result in milliseconds regardless of volume
  • The agent gets a deterministic answer
  • Many agents in parallel, zero conflict
// Simulation: traversal vs. facet lookup
// Click ▶ Traversal BFS or ⚡ Facet lookup to see the difference
07 / Solution

Three-tier architecture

The graph remains the source of truth. Precomputed indexes and projections are the interfaces agents actually use — never the raw graph.

Tier 2
Facets
Precomputed indexes — O(1) response
∩
Roaring bitmaps — each facet value maps to a bitmap of matching entity IDs. Intersection = logical AND.
#
Exact counts — how many results per facet, precomputed, no scan.
⚡
Multi-criteria at once — sector=tech AND budget>50k AND city=BRU in one operation.
∥
Parallel reads — many agents, zero read locks, deterministic results.
→
Tier 3
Projections
Precomputed subgraphs per agent role
◈
Dedicated view per agent — a legal agent only sees Law/Article/Decision nodes. Not the full graph.
⊂
Context reduction — fewer tokens, less noise, sharper reasoning.
↻
Automatic sync — projections recompute when the source graph changes.
🔒
Isolation and security — each agent only accesses its authorized knowledge scope.

Example: one enterprise graph, three different projections depending on which agent accesses it.

Sales agent
// projection: sales_view
Contact Account Opportunity Activity Contract Statute article Invoice
Legal agent
// projection: legal_view
Contact Account Opportunity Contract Statute article Clause Case law
Finance agent
// projection: finance_view
Contact Account Opportunity Contract Invoice Payment Case law
08 / Context

Why Neo4j, ArangoDB, and others are not enough

Existing graph systems are built for human developers who write queries. AI agents face radically different constraints.

⚠ The impedance problem

A developer writes a Cypher or SPARQL query once, tests it, optimizes it. An agent must dynamically produce a correct query on every call, never get syntax wrong, never trigger an unintentional full scan, and stop within a predictable time budget. That is not the same problem.

O(V+E) Raw traversal complexity on an unindexed graph
< 3ms Precomputed facet lookup on Wikidata (141 critical queries, GraFa 2020)
×10–100 Slowdown of graph DB vs native index (Neo4j vs SQL benchmark)
✓ mindBrain principle

Agents never touch the raw graph. They query deterministic interfaces — facet indexes for multi-criteria lookups, projections for relational reasoning — and receive answers bounded in time and size.

09 / Stakes

A company is a sum of disjoint ontologies

Real customer value does not live in any single silo. It emerges at their intersection — and today that intersection only lives in a senior rep's head.

ERP

Contracts & finance

  • Signed contracts
  • Billing and payments
  • Deliveries and inventory
  • Pricing terms
// ignore: relationship history, tickets
CRM

Sales pipeline

  • Contacts and accounts
  • Opportunities and pipeline
  • Sales activities
  • Revenue forecasts
// ignore: invoices, incidents, clauses
Support

Customer relationship

  • Tickets and incidents
  • SLAs and commitments
  • Satisfaction and NPS
  • Incident history
// ignore: opportunities, contracts, payments
∩ Intersection — what is missing everywhere

"This customer has 3 active contracts, 2 P1 tickets open for 15 days, a renewal in 60 days, and a rep who has not touched base in 45 days." — That sentence exists in none of the three systems.

✔ The mindBrain thesis

Salesforce changes its API unilaterally. SAP and Oracle charge you to extract your own data. Odoo is open source but its ontology is still encoded in structures you do not control. Data sovereignty starts with ontology sovereignty.

WHAT CHANGES IN PRACTICE

What changes in practice

Without mindBrain

  • – re-read long notes on every session start
  • – reconstruct project structure from prose
  • – guess at dependencies it cannot see
  • – operate without knowing what it does not know

With mindBrain and GhostCrab

  • → ask what is blocked and get a structured answer
  • → traverse dependencies without scanning everything
  • → receive compact, pre-ranked working context for the current task
  • → operate across sessions without starting from zero

The agent's reasoning capability does not change. What changes is the quality of the environment it reasons in.

THE ONTOLOGY BENEATH EVERYTHING

The ontology beneath everything

mindBrain is not just a storage layer. It holds a domain ontology — the typed model of what exists in a project,
what relationships are valid, and what rules govern transitions.

This is what separates mindBrain from a generic database or vector store. A vector store preserves proximity.
A graph database preserves connections. mindBrain preserves meaning — the structured intent behind a domain,
expressed as typed entities, directed relationships, and constraint rules.

An agent working on a CRM project does not need to know about HR workflows. An agent managing an onboarding case
does not need access to legal contracts from other clients. mindBrain's ontology defines the boundary of what is relevant — and GhostCrab enforces it.

Ontologies can also build bridges. mindBrain supports cross-domain connection ontologies — structured mappings
that link entities across otherwise siloed models: ERP, CRM, HR, project management, helpdesk. This is where agents
can answer questions that no single system can today.

Cross-domain example

Which customers with contracts over €50k have an open P1 support ticket unresolved for 15 days, a contract expiring in 30 days, and no sales contact in the last 45 days?

Data sovereignty starts with ontology sovereignty.

You do not control your data if you cannot control its structure.

IN SHORT

In short

mindBrain Personal

Structured agentic database — SQLite engine — for individuals and small teams

mindBrain Professional

Structured agentic database — PostgreSQL engine — for teams and enterprise

GhostCrab

MCP server — the agentic interface to mindBrain

Your agent

Reasoning, conversation, intent — unchanged

GhostCrab gives your agent a workspace. mindBrain gives that workspace memory, structure, and meaning.

Open source · Self-hostable · MCP-compatible

Install on GitHub See use cases Read the architecture
INSTALL ON GITHUB EXPLORE USE CASES READ THE ARCHITECTURE
GhostCrab mascot

Hello, I'm GhostCrab — your favorite MCP server.