Skip to content
AINew: 5-layer quality architecture for cited answers
Subsum•ioLEGAL INTELLIGENCE
Pricing
Subsum•ioLEGAL INTELLIGENCE

AI legal software that never forgets — the firm brain for lawyers in Europe.

Your data, your keys. Self-hosted on your hardware or our EU cloud with DPA — GDPR-ready, end-to-end encrypted, zero training on your data.

Platform

  • Overview
  • SuperBrain
  • Features
  • Security
  • WhatsApp Copilot
  • Pricing
  • Download

Solutions

  • For Individuals
  • For Legal Professionals
  • For Law Firms
  • For Solo Lawyers
  • For In-House
  • For Mid-Sized Firms
  • For Tax Advisors

Resources

  • Documentation
  • Blog
  • Benchmark
  • Partner Program
  • Dashboard

Company

  • About
  • Contact
  • Imprint

Legal

  • Terms of service
  • Privacy

© 2026 Subsumio · AI legal software & legal intelligence for law firms

EU cloud or self-hosted · GDPR-ready · AVV included · § 203 StGB compliant

The Subsumio SuperBrain

Not just AI.An artificial brain for your firm — regenerating every night.

Other AI tools remember chat histories. Subsumio understands your cases: every night, the SuperBrain consolidates them into a living legal memory — like the human hippocampus during sleep. Five quality layers, contradiction probing, citations. Zero hallucinations.

SuperBrain
Extract facts
Synthesize
Consolidate
Engram maturation
Reconsolidation
Embeddings
Detect patterns
Contradiction probe
Grade takes
Calibration
0

Nightly brain phases

Consolidation, engram maturation, reconsolidation, contradiction probing

0

Correction layers

Extraction → Synthesis → Verdict → Probe → Ensemble

0.00

LEXam score

DeepSeek V3.2 — better than GPT-4o

0.0%

Recall@8

LongMemEval, 500 questions

The Problem

Prompt AI: asks, answers, forgets

You ask a question. The model answers. Then it forgets everything. Other tools store chat histories and call it 'Memory'. Subsumio builds a knowledge graph — not chat storage, but understanding. Every prompt in other tools is isolated — no memory, no learning, no understanding. Hallucinations are fought at the prompt level: with hope. The model is judge and defendant at once.

Prompt

You ask GPT-5 or Claude

Answer

Model generates — without memory

Forget

Everything is discarded. Next prompt = zero context.

✕Chat-history storage instead of a knowledge graph — chats are stored, not understood
✕No nightly consolidation — knowledge doesn't accumulate
✕Hallucinations fought at prompt-level (hope-based)
✕No contradiction detection between answers
✕Premium pricing for a single front-end model with no memory
✕No legal quality control — the model is judge and defendant
The Solution

SuperBrain: learns, consolidates, understands

Subsumio builds a knowledge graph from your documents every night. Facts are extracted, synthesized, probed for contradictions, and validated by an ensemble judge that surpasses human lawyers. The result is not a chat history. It's a long-term legal memory that gets sharper every night — not just bigger, but more understanding. The longer you use it, the better the SuperBrain understands your firm.

01

Ingest

Documents, emails, matters, conversations

02

Process

5 layers: Extraction → Synthesis → Verdict → Probe → Ensemble

03

Store

Persistent knowledge graph with facts, takes & relationships

04

Answer

Every answer from the graph — with citations

Every night: 29 phases · 5 layers · 1 knowledge graph

The 5-Layer Architecture

Every take passes through five quality layers before entering your knowledge graph. Verified LEXam scores per model. This is the architectural moat no competitor can cross.

00

Layer 0 — Fact Extraction

Specialized extraction models structure facts from conversations, documents and emails. No legal reasoning — just structuring. Every night, new facts are written to the brain.

119 languages · Fully automatic · Source references at document level
01

Layer 1 — Legal Synthesis

DeepSeek V3.2 synthesizes facts into takes with legal reasoning. LEXam score 57.42 — better than GPT-4o (56.93), on par with GPT-4.1 (57.50). Open-weight, fully transparent.

DeepSeek V3.2 · LEXam 57.42 · Open-weight · Fully traceable
02

Layer 2 — Verdict

A reasoning model evaluates every single take: accepted, rejected, or sent back for revision. Bad takes die before they reach the graph. LEXam score 56.53 — nearly identical to GPT-4.1.

Model: DeepSeek V3.2-Reasoner · LEXam 56.53 · Evidence scoring with confidence
03

Layer 3 — Contradiction Probe (tiered)

Chunk pairs are nightly checked for logical contradictions. 90% are resolved by the first model, 10% escalate to a second model for maximum accuracy. Contradictions are flagged — not hidden.

DeepSeek V3.2 (initial) → Claude Haiku (10% escalation) · Cross-reference audit
04

Layer 4 — Ensemble Legal Gate

An ensemble of two open-weight models evaluates final legal quality. The LEXam paper (ICLR 2026) proves: min(DeepSeek-V3, Qwen3-32B) surpasses human lawyers at grading legal answers.

Ensemble: min(DeepSeek V3, Qwen3-32B) · Surpasses human lawyers · LEXam-validated (ICLR 2026)

5 layers. 29 phases. A knowledge graph that grows every night — and gets richer through reconsolidation.

Nightly brain regeneration

While you sleep, the SuperBrain runs 29 phases — just like the human brain repeats, strengthens, and rewires memories during sleep. Facts become knowledge, takes become understanding, contradictions become clear signals. New: Engram Maturation and Reconsolidation — inspired by neuroscience. All automatic, all monitored, all logged.

01 / 10

Extract facts

Conversations, emails, documents → structured facts with source references

02 / 10

Synthesize

Facts → brain pages with legal reasoning and citations

03 / 10

Consolidate

Cluster facts → promote takes → remove duplicates

04 / 10

Engram maturation

New memories start 'silent' and mature over 7 days via a sigmoid curve — like the human hippocampus. Implicit knowledge becomes explicit understanding.

05 / 10

Reconsolidation

When a memory is retrieved, it becomes 'labile' for 60 minutes — new nuances can merge in, like in the human brain. Knowledge gets richer, not just bigger.

06 / 10

Embeddings

Chunks → vector space for semantic search across all matters

07 / 10

Detect patterns

Cross-take themes, connections and recurring risks

08 / 10

Contradiction probe

Logical conflicts between takes are flagged — not hidden

09 / 10

Grade takes

Legal quality scoring with LEXam-validated models

10 / 10

Calibration

Track model quality over time · Detect drift · Adjust

Fully automatic · Nightly · Monitored

Other legal AI vs. SuperBrain

Other legal AI uses multi-model routing per task. We use multi-model quality layers per take — with a knowledge graph that consolidates every night. That's not a feature difference. That's an architectural paradigm shift.

Architecture

Other legal AI

✕Multi-model per task (Harvey) or single-model

Subsumio SuperBrain

5-layer quality pipeline per take

Hallucination control

Other legal AI

✕Prompt-level (hope-based)

Subsumio SuperBrain

Architectural (5 correction layers + ensemble judge)

Memory / Recall

Other legal AI

✕Chat-history storage or announced

Subsumio SuperBrain

Knowledge graph with nightly consolidation — since day 1

Knowledge graph

Other legal AI

✕None — every prompt is isolated

Subsumio SuperBrain

Persistent graph with facts, takes, relationships

Nightly consolidation

Other legal AI

✕Not present

Subsumio SuperBrain

29 phases: consolidation, engram maturation, reconsolidation, contradiction probing

Engram maturation

Other legal AI

✕Not present — no concept of memory maturation

Subsumio SuperBrain

Sigmoid maturation curve over 7 days — silent → implicit → explicit

Reconsolidation

Other legal AI

✕Not present — memories are static

Subsumio SuperBrain

Labile window on retrieval — knowledge gets richer through nuances

Legal benchmark

Other legal AI

✕LAB (Harvey) or none

Subsumio SuperBrain

LEXam (ICLR 2026) — 340 exams, 7,537 questions

Multilingual

Other legal AI

✕50-95 languages

Subsumio SuperBrain

119 languages (Qwen3.5 engine)

Pricing model

Other legal AI

✕Premium per-seat — you pay for their API costs

Subsumio SuperBrain

Fair pricing — no hidden API costs

DACH law fine-tuning

Other legal AI

✕None (general models)

Subsumio SuperBrain

Subsumio Legal-32B (proprietary, law-corpus)

GDPR

Other legal AI

✕US cloud (Anthropic/OpenAI)

Subsumio SuperBrain

EU cloud or self-hosted
FeatureOther legal AISubsumio SuperBrain
Architecture
✕Multi-model per task (Harvey) or single-model
5-layer quality pipeline per take
Hallucination control
✕Prompt-level (hope-based)
Architectural (5 correction layers + ensemble judge)
Memory / Recall
✕Chat-history storage or announced
Knowledge graph with nightly consolidation — since day 1
Knowledge graph
✕None — every prompt is isolated
Persistent graph with facts, takes, relationships
Nightly consolidation
✕Not present
29 phases: consolidation, engram maturation, reconsolidation, contradiction probing
Engram maturation
✕Not present — no concept of memory maturation
Sigmoid maturation curve over 7 days — silent → implicit → explicit
Reconsolidation
✕Not present — memories are static
Labile window on retrieval — knowledge gets richer through nuances
Legal benchmark
✕LAB (Harvey) or none
LEXam (ICLR 2026) — 340 exams, 7,537 questions
Multilingual
✕50-95 languages
119 languages (Qwen3.5 engine)
Pricing model
✕Premium per-seat — you pay for their API costs
Fair pricing — no hidden API costs
DACH law fine-tuning
✕None (general models)
Subsumio Legal-32B (proprietary, law-corpus)
GDPR
✕US cloud (Anthropic/OpenAI)
EU cloud or self-hosted
Proprietary Moat

Subsumio Legal Engine

We fine-tune open-weight models on our own law-corpus: BGB, ZPO, HGB, StGB, AO, ABGB, OR — the entire DACH legal code. The result is a proprietary model no competitor has.

law-corpus40+ statutes

BGB, ZPO, HGB, StGB, AO, EStG, UStG, ABGB, OR, GDPR — DE/AT/CH/EU

Base modelQwen3-32B

119 languages · Open-weight · LoRA fine-tuning with Unsloth

Fine-tuningLoRA

Parameter-efficient fine-tuning · Open-weight · No vendor lock-in

DeploymentSelf-hosted

EU cloud or fully on-premise · No API dependency

Forecast

Forecast: +5–10 LEXam points over base Qwen3-32B (40 → ~45–50). Ensembled with DeepSeek V3.2: quality at GPT-4.1 level (57.50) — with a model that understands DACH law from the ground up.

What you can do with the SuperBrain

From the first matter to complex multi-party cases — the SuperBrain works in every case. Here are the concrete use cases law firms use daily.

Legal research with citations

Questions about a case? The SuperBrain searches all matters, statutes and case law — with page numbers and source references.

Contract analysis & red-lining

Contracts are automatically analyzed, risks flagged, and amendment suggestions generated — with BGB/ABGB references.

Deadline tracking

Automatic deadline calculation per ZPO, BGB, ABGB. The SuperBrain monitors and alerts — before deadlines expire.

Knowledge management

Every matter, every correspondence becomes a knowledge graph. Never solve the same problem twice.

Conflict checking

Cross-reference audit finds conflicts of interest before they become a problem — across all clients.

Client reporting

Automatic summaries, status reports and cost estimates — straight from the knowledge graph.

Explore all features

Your data. Your keys. Your jurisdiction.

The SuperBrain processes everything inside your isolated environment. No training on your data. No data to third parties. Confidentiality by architecture — not by promise.

Per-matter isolation

Every case is logically separated. Client confidentiality stays confidential.

No training on your data

Models never learn from your documents. Architecturally guaranteed.

EU cloud or self-hosted

Hetzner Falkenstein (DE) or fully on-premise.

Auditable

Every take, every verdict, every calibration is traceably logged.

Compliance by architecture

The SuperBrain wasn't made GDPR-compliant after the fact. Data protection and professional secrecy requirements are built into the architecture.

GDPR

EU cloud, no data to third parties

§ 203 StGB

Professional secrecy by design

GoBD

Procedure documentation integrated

BRAO

Law firm compliant from day one

Integrated into your firm workflow

beASecure legal communication
DATEVTax & accounting
DocuSignElectronic signatures
WhatsAppClient communication
OutlookEmail & calendar
ADVOKATLaw firm software import
Read security details

Frequently asked questions

Everything you need to know about the SuperBrain, the Dream Cycle and the 5-layer architecture.

The SuperBrain is the AI engine behind Subsumio — an artificial legal brain that regenerates every night. It builds a knowledge graph from your documents with 29 automated phases and 5 quality layers. Every answer comes from your matters — with citations, not from a generic model.

The Dream Cycle is the SuperBrain's nightly consolidation loop. It mirrors what happens in the human brain during sleep: information is replayed, strengthened, connected to existing knowledge, and conflicting or stale fragments are pruned. 29 phases, fully automatic, everything logged. New: Engram Maturation (memories mature over 7 days) and Reconsolidation (knowledge gets richer on retrieval).

LEXam (ICLR 2026) is a legal benchmark with 340 exams and 7,537 questions. DeepSeek V3.2 scores 57.42 — better than GPT-4o (56.93) and on par with GPT-4.1 (57.50). The SuperBrain uses these verified scores for every model decision.

Yes. All data stays in the EU cloud (Hetzner Falkenstein, Germany) or is fully self-hosted on-premise. No training on your data, no data to third parties. Per-matter isolation ensures professional secrecy under § 203 StGB.

Architecturally, not hope-based. Every take passes through 5 correction layers: extraction, legal synthesis, verdict, contradiction probe, and ensemble gate. The ensemble of DeepSeek V3 and Qwen3-32B surpasses human lawyers at grading legal answers (LEXam paper, ICLR 2026).

Harvey uses multi-model routing per task (which model for which job). Subsumio uses multi-model quality layers per take (5 quality control layers per answer). Subsumio has a persistent knowledge graph with nightly consolidation, contradiction detection, and fine-tuning on DACH law. Subsumio is EU cloud or self-hosted — Harvey runs on US cloud.

ChatGPT and Claude store chat histories and preferences — they 'remember'. The SuperBrain 'understands': it builds a structured knowledge graph from your documents every night, checks for contradictions, grades legal quality with LEXam-validated models, and gets sharper with every night. ChatGPT Memory is like a sticky note. The SuperBrain is like an associate who goes through your files every night and understands more by morning than the evening before.

Yes. Subsumio can run fully on-premise in your own infrastructure. No API dependency, no data leaves your network. Ideal for firms with strict compliance requirements.

119 languages via the Qwen3.5 engine — including German (DE, AT, CH), English, French, Italian, Polish, Turkish and Arabic. Perfect for international firms.

We fine-tune open-weight models (Qwen3-32B) on our law-corpus: BGB, ZPO, HGB, StGB, AO, ABGB, OR — the entire DACH legal system. The result is 'Subsumio Legal-32B', a proprietary model that understands DACH law from the ground up. No competitor has this.

The SuperBrain is included in all Subsumio plans. You pay per user — no hidden API costs, no token calculations. 14 days free, no credit card. See our pricing page for details.

See pricing plans

Experience the SuperBrain

Start today. Your legal brain grows from the first matter — every night, automatically, gaining understanding. 14 days free, no credit card.

Start 14-day free trialNo credit card · Learns every night
SuperBrain — The AI Engine Behind Subsumio | Legal AI with 5-Layer Architecture — Subsumio