HEINRICH
HEINRICH · HHI · Structurally New · Built to be Agentic

We Broke
The Mold

A structurally new architecture, built to be Agentic. Heinrich is the Harmonic Mind — where every concept lives at its own coordinated location, every connection traces a path, and every answer is known before it is given. Designed to run on the hardware you already own.

Request Early AccessRead the ArchitectureSee HeinrichEnvironmental Impact
A Language Model
stateless · GPU
01
Tokenize
Text broken into subword tokens. Full prior context re-tokenized every call.
↓
02
Embed + Position
Each token mapped to a high-dimensional vector with positional encoding.
↓
03
Transformer Block × N layers
Self-AttentionQ · K · V
Feed-ForwardMLP / NonLin
Repeated 32–96 layers. Every token sees every prior token through attention.
↓
04
LM Head + Softmax
Final hidden state projected to the vocabulary, normalized to a probability distribution.
↓
05
Sample One Token
Greedy / top-k / nucleus sampling picks the next token from the distribution.
↺ Loop — full forward pass repeated for every single token
Heinrich · Harmonic Mind
stateful · CPU
01
Goal Intake
Request captured as structured objective, constraints, permissions, success criteria.
↓
02
Knowledge State Lookup
Concept queried in index. Heinrich asks itself first: Known / Partial / Uncertain / Stale.
↓
03a · if Known
Harmonic Retrieval
Concept accessed at its coordinated location in the substrate. No generation.
03b · if Gap
Research Cortex
Gap detected → source under permission → staged as evidence for review.
↓ merge ↓
04
Cognitive Layer Routing
Concept routed through the appropriate cognitive layer — Causal, Quantity, and others.
↓
05
Hebbian Association Prior
Co-activated concepts strengthen retrieval paths. Bounded, decaying, reviewable.
↓
06
Plan + Approval Gate
Goal decomposes into tasks with criteria. Sensitive actions pause for user approval.
↓
07
Supervised Agent Execution
Worker agents run the plan. Blockers surface. Checkpoints persist throughout.
↓
08
Verification Loop
Work checked against goal criteria before close. Failures route back, not forward.
↓
09
Proof Receipt + Memory Update
Checkpoint, evidence bundle, and audit row emitted. Hebbian links strengthened. Knowledge State updated.
The Simple Distinction

Two different categories of thing.

People often compare LLMs and Heinrich as if they are the same kind of product at different quality levels. They are not. The difference is architectural — like comparing a calculator to a library.

○ What an LLM is

A probabilistic text generator.

A large language model is a trained statistical network that, given a prompt, predicts the most likely sequence of tokens to follow. It has no goals. No persistent memory. No native concept of confidence. It does not know what it did last time. It produces text. That is the work.

This is not a criticism — it is the definition. LLMs are extraordinary at the thing they actually do. They are not built to do the rest.

● What Heinrich is

A Harmonic Mind.

Heinrich is a structured cognitive architecture built by EMPHOS Group. Concepts live as addressable units in a harmonic substrate. Memory is biological — concepts that activate together strengthen retrieval paths. A Knowledge State tracks confidence, evidence, and answer-readiness. Research closes its own gaps under permission. Workflows produce proof receipts.

It does not predict the next token. It does not use a language model. It runs on the CPU you already own.

An LLM generates text. Heinrich is a mind. Asking whether one is "better than" the other is the wrong question — they do different things.
Same Workload, Measured

100 million agentic tasks a day. What each one costs to run.

Put all four on the same job: 100 million agentic tasks a day, about what OpenAI serves now, each expanding into roughly 25 reasoning and tool steps. Same work, very different infrastructure. Shorter bars are better.

Servers required25 to 50x fewer
HeinrichCPU
800–1,400
ClaudeGPU
30k–45k
GPTGPU
37.5k–55k
GeminiTPU
25k–37.5k
Continuous power (MW)30 to 75x less
HeinrichCPU
6–10
ClaudeGPU
375–625
GPTGPU
450–750
GeminiTPU
300–500
Daily water (litres)~10 to 500x less
HeinrichCPU
50k–100k
ClaudeGPU
2–5M
GPTGPU
2.5–50M
GeminiTPU
650k
One line none of them can cross. Heinrich ships to the edge at under 50 MB, on-device and offline. Claude and GPT are cloud-only; Gemini reaches it only via Nano at ~1.8 GB.
Sources. Frontier figures come from public disclosures and independent studies: OpenAI 0.34 Wh per query, Google 0.24 Wh per Gemini query, ~2.5 billion queries a day. Heinrich’s are an EMPHOS projection from its CPU WaveField architecture, every assumption set against Heinrich, so the real gap should be wider. Pre-production projection.
One to two orders of magnitude less hardware, power, and water, all for the same work, because an answer is an address, not a GPU network run per token.
Capability Matrix

At a glance, across the field.

A direct comparison of what each system is structurally built to do. Not a feature war — an architectural map. Most of the gaps are not bugs in the LLMs. They are categories of behavior an LLM is simply not built to provide.

CapabilityClaudeGPT-5GeminiQwenDeepSeekGrokHeinrich
Structural Architecture
Statistical next-token prediction
Structured concept substrate
Addressable concept retrieval
Knowledge & Memory
Persistent memory across sessions
Knowledge confidence routing
Known / Partial / Uncertain / Stale
Hebbian co-activation memory
Agentic Operation
Native goal planning
Supervised execution loop
Verification before close
Approval gates on sensitive actions
Tool / action contracts
Accountability
Proof receipts (per-step)
Persistent audit trail
Reviewable evidence bundles
Rollback awareness
Hardware & Operation
CPU-only execution
No GPU required
Runs locally on the device
Sub-second response on commodity hardware
Stateful across sessions
Native · built into the architecture
Yes · supported
Partial · possible via configuration or wrapper
—Not native · not part of the system itself
Heinrich operates in the empty spaces. Not because LLMs are broken — because intelligence built around goal completion, governed memory, and proof is a different category of product than a language model is built to be.
Side by Side

The architectural gap, line by line.

Where the two diverge — not in marketing copy, but in what each actually does at runtime.

Dimension
Typical LLM
Heinrich
Core mechanism
Statistical next-token prediction.
Structured concept substrate with explicit retrieval — no token prediction.
Knowledge representation
Distributed across all parameters. Implicit.
244M addressable concepts. Explicit. Each one named, located, and queryable.
Memory
Stateless between calls. Context resent every turn.
Hebbian co-activation. Biological. Bounded, decaying, reviewable.
Knowledge confidence
No native sense of certainty. Confident-sounding regardless.
Knowledge State Index: Known / Partial / Uncertain / Contradicted / Stale — before answering.
Goal awareness
None. Each call is an isolated completion of a prompt.
Goal orchestration with milestones, criteria, approval gates, checkpoints.
Research
Not native. Falls to the wrapper.
Inner Mind Research Cortex: identifies gaps, fetches under permission, stages evidence, gates promotion.
Verification
None native. The user verifies, or no one does.
Verification loops check work against the goal — before close.
Accountability
No persistent reasoning trail. Output only.
Proof receipts: checkpoints, evidence bundles, audit rows. First-class output.
Hardware
GPU clusters in remote data centers.
CPU. Local. Sub-second on hardware you already own.
Posture
Reactive. Waits for the next prompt.
Agentic. Plans, acts, surfaces blockers, asks for approval where needed.
None of these capabilities come from making the language model bigger. They come from building a different kind of system altogether.
The "Bigger Model" Argument

Scale does not fix what is structurally missing.

It is tempting to assume the next, larger model will close the gap. Heinrich is built on a different reading of the evidence: certain capabilities are not emergent from scale. They are emergent from architecture.

Myth 01

"A larger LLM will remember."

Larger context windows help. They do not add persistent memory across sessions, durable storage of decisions, or a clean separation between what is fact and what was merely said in conversation. Those are architectural — not size-driven.

Heinrich: Hebbian co-activation memory, bounded and reviewable, with verified knowledge clearly separated from unverified context.
Myth 02

"A larger LLM will stop hallucinating."

Frontier models still confidently produce incorrect output. The root cause is that the model has no native concept of knowing whether it knows. Confidence and correctness are not the same signal, and scaling alone does not align them.

Heinrich: the Knowledge State Index decides answer-readiness before generation begins. If the concept is not known, Heinrich routes to research or asks — it does not invent.
Myth 03

"A larger LLM will plan for me."

Models produce plan-shaped text. They do not natively execute plans, track state across execution, recover from failures, or surface blockers. That is the job of a runtime, not a generator.

Heinrich: goal orchestration with approval gates, supervised execution loops, persistent checkpoints, and rollback awareness.
Myth 04

"A larger LLM will run on my laptop."

The trajectory is the opposite. Each generation requires more compute, more memory bandwidth, more energy. The largest models will not fit on consumer hardware in any reasonable horizon.

Heinrich: CPU-only by design. Sub-second response on the commercial benchmark path. No GPU. No data center round-trip.
Scaling the model is the wrong axis for these problems. The right axis is system design.

"A language model predicts the next word. A Harmonic Mind knows the concept — and decides whether it should answer at all."

Heinrich design principle · EMPHOS Group

Sometimes smaller is better - EMPHOS

Heinrich is the EMPHOS bet on a different architectural lineage — the Harmonic Mind. CPU-only. Knowledge-aware. Proof-first. If that distinction matters to you, get in touch.

Request Early AccessRead the ArchitectureSee HeinrichEnvironmental Impact