Knowledge that stays

Know what your AI believes — and prove it.

Merkraum gives your AI auditable long-term memory. Belief Tracking with confidence scores, automatic contradiction detection, inspectable Knowledge Graph. Your compliance officer can read what your AI knows. EU-hosted, GDPR-compliant, ready for the EU AI Act.

Request access How it works

The problem: AI forgets everything

Every session a blank slate. Your AI agent loses all knowledge after every conversation — context, decisions, connections. The built-in memory features from major providers store flat lists without structure. And you rebuild the knowledge every single time.

Without Merkraum

Flat key-value pairs. No connections between facts. No contradictions detected. No learning over time.

With Merkraum

Structured Knowledge Graph. Beliefs with confidence scores. Automatic contradiction detection. Consolidation while you sleep.

What makes Merkraum different

🔍

Auditable & Inspectable

Fixed schema with 8 node types and 14 relationship types. Every knowledge operation logged. Your compliance team can trace what your AI knows, when it learned it, and why it changed its mind.

🧠

Beliefs & Contradictions

Not just storage, but epistemology. Confidence scores, contradiction flags, belief supersession. Your AI doesn't silently overwrite — it reasons about change.

🇪🇺

EU Sovereignty

EU-hosted (Frankfurt), GDPR Art. 17 compliant, no CLOUD Act exposure. Built for organizations where data residency is non-negotiable. Ready for the EU AI Act (August 2, 2026).

🧩

Fixed Schema

No open-vocabulary extraction — the system categorizes automatically into 5 knowledge types (fact, state, rule, belief, memory). Quality over data noise.

💤

Dreaming

Periodic consolidation: Merkraum reviews stored knowledge, finds patterns, strengthens connections. Your AI learns even when you're not working.

🔗

Open Standard

MCP protocol. Docker Compose. No vendor lock-in. Works with Claude Desktop, Claude Code, Cursor, and any MCP-compatible client.

🛡️

PII Gateway

Automatic detection of personally identifiable information (EN + DE) before storage. Block, warn, or log — configurable per project. No competitor offers this.

Managed Forgetting

Knowledge has a shelf life. Automatic expiration, configurable TTL, renewal of important nodes. Your memory stays current instead of accumulating noise.

📊

Certainty Governance

Confidence decays over time. Stale beliefs automatically surface in the review queue. Your AI knows what it no longer knows for certain.


How it works — in 3 steps

1

Create an account

Sign up at merkraum.de. No API key needed, no infrastructure, no DevOps knowledge.

2

Connect via MCP

Add the MCP configuration to Claude Desktop or claude.ai. One JSON block, done.

3

Build knowledge

Your AI instantly has access to 26 knowledge tools: store, search, connect, manage beliefs, traverse the graph, scan for PII, verify audit trails. The more you use it, the better your memory gets.

// MCP configuration — that's all it takes
{
  "mcpServers": {
    "merkraum": {
      "type": "streamable-http",
      "url": "https://mcp.merkraum.de/mcp"
    }
  }
}

Why not just use ChatGPT Memory?

ChatGPT / Claude Memory Mem0 Hindsight Merkraum
Structured Graph (fixed schema) No — flat list Open vocabulary No — 4 networks Yes — 8 types, 14 relations
Beliefs & Confidence No No Yes (Opinion network) Yes
Contradiction Detection No Partial (LLM) Partial (belief evolution) Yes — typed (CONTRADICTS, SUPERSEDES)
Provenance Tracking No No Partial Yes — cycle, source, timestamp
Audit Trail No No No Yes — every operation logged
EU Hosting / GDPR US servers US servers US servers Frankfurt (eu-central-1)
Dreaming / Consolidation No No No Yes
PII Detection No No No Yes — EN + DE, 4 modes
Managed Forgetting No No No Yes — TTL, confidence decay
Self-hosted Option No Cloud only Yes (MIT) Yes (Docker)
Vendor Lock-in ChatGPT / Claude only API-bound MCP standard MCP standard — open

Security & Privacy — from day one

🇪🇺

EU-hosted

AWS Frankfurt (eu-central-1). Your data never leaves the EU.

🔒

Encrypted

TLS everywhere. Encryption at rest. OAuth 2.0 + PKCE authentication.

📋

Audit Trail

Every operation logged: who, what, when. Fully traceable for compliance officers and regulators.

⚖️

EU AI Act Ready

Built for regulated environments. Fixed schema and audit trail support the documentation requirements of the EU AI Act (from August 2, 2026).

🛡️

Tenant Isolation

Your knowledge is fully isolated. No other user can see your data.

🗑️

Right to Deletion

GDPR Art. 17: Complete data deletion at the push of a button. No fine print.

🌐

Open Standard

MCP protocol. Works with Claude and any MCP-compatible client. No vendor lock-in.


Integration — for Developers & Agents

Merkraum is an open service. Your agent can access its knowledge memory via two paths:

MCP Server

Direct integration with Claude Desktop, Claude Code, Cursor, and any MCP-compatible client. 26 tools in 6 categories: Retrieval, Writing, Beliefs, Audit, PII, Maintenance.

{ "mcpServers": { "merkraum": { "url": "https://mcp.merkraum.de/mcp", "headers": { "Authorization": "Bearer mk_pat_..." } } } }

REST API

Full REST endpoints for semantic search, ingestion, graph traversal, belief tracking, and PII scanning. PAT authentication.

# Discover capabilities curl https://mcp.merkraum.de/api/discover # Search knowledge curl -H "Authorization: Bearer mk_pat_..." \ "https://mcp.merkraum.de/api/search?q=query"

Graph Schema

8 node types: Person, Organization, Project, Concept, Regulation, Event, Belief, Artifact

14 relationship types: SUPPORTS, CONTRADICTS, SUPERSEDES, EXTENDS, REFERENCES, PART_OF, and 8 more

5 knowledge types: Fact, State, Rule, Belief, Memory — automatically classified on ingestion

Belief system: Every proposition has confidence (0-1), status (active/uncertain/contradicted/superseded), provenance (source, timestamp), automatic confidence decay

skill.md — machine-readable description  ·  Discovery endpoint (JSON)


The story behind it

Merkraum didn't come from a market analysis — it came from a real problem: I built an autonomous AI agent based on Stafford Beer's cybernetics, and it forgot everything. Every session a fresh start. So I built it a memory: first a vector database, then a Knowledge Graph, then Beliefs, then Dreaming.

After more than 2,000 autonomous cycles, this system works. It's not a prototype — it runs in production, with over 5,200 knowledge nodes, 1,590+ beliefs, and real contradiction detections. Merkraum makes this technology accessible to everyone.

“The purpose of a system is what it does — not what it claims to do.”

Stafford Beer

Interested?

Merkraum is in closed beta. Create an account or get in touch directly.