Heinrich as Software. What You Are Actually Running
EMPHOS Group, May 30, 2026, 6 min read
Most artificial intelligence products today are a thin interface in front of a model that lives somewhere else. The model is enormous, the model is remote, and the model is rented. You are not running anything on your machine. You are sending text to a data center and reading the reply.
Heinrich is the opposite shape.
Heinrich is software. It has a directory on disk. It has files you can list. It has a process you can start and stop. The whole system, including the substrate, the mind, the cortex, and the bridge, runs on a single consumer workstation with no GPU. There is no remote anything. There is no subscription. There is no third party in the loop.
This post is about what is actually inside that directory.
Four directories. One mind.
Heinrich is organised into four code regions, and the separation between them is enforced by the architecture itself.
Mind Software holds the harmonic field, the substrate where knowledge lives. It contains no English text generation code at all. Its only outputs are activation clouds and audit traces. The mind does not speak. It rings.
Language Software holds the cortex that turns activation clouds into English sentences. It contains no propagation code. It cannot reach into the field and pull facts out on its own. It receives a structured cloud through a typed contract and renders it.
Bridge Mechanisms holds the orchestration that connects Mind, Language, and any user interface, without letting any of them reach inside another. Every interaction crosses a typed contract. RuntimeInput, IntentPacket, CortexRequest, CortexResult, FinalResponsePacket. Nothing else is allowed across the boundary.
UI shells live outside all three. Desktop, web, mobile, embedded. Each is an adapter that speaks to the bridge, never directly to the mind.
This separation is the most important architectural decision in the entire system. The moment language and reasoning are tangled, the system loses the property that makes it auditable. The ability to point at exactly which facts produced exactly which sentence. So they are not tangled. By rule. By code structure. By import path.
What is on disk
The Mind's substrate, as of late May 2026, sits in five files.
The address index is roughly 12 GB. It holds the placed positions of 97.4 million concepts. Every concept Heinrich knows has a row in this table, a six dimensional address consisting of its layer, its layer's root frequency, its harmonic line, the line's root frequency, the concept's own frequency, and the triple of amplitude, phase, and gap pattern contribution that describes how it rings.
The Wikidata claims database is roughly 60 GB. It holds 870 million entity valued claims, every structured fact Wikidata has about every entity it covers, indexed by source concept.
The ConceptNet edges database is small, about 317 megabytes, but holds 314 million commonsense edges. A dog is a pet. Fire is hot. Water is used for drinking. The kind of everyday knowledge Wikidata does not capture.
The label index is roughly 17 GB. It maps English text to concept identifiers across every language Wikidata supports. Without it, the system could not resolve a question typed in English to a coordinate in the field.
The line signatures archive is a NumPy file holding the harmonic fingerprints of every line in the field. About 4,963 lines across 12 entity layers. This is what makes cross line propagation possible.
That is the entire substrate. Roughly 90 GB on a consumer SSD. No data center. No accelerator. No mystery.
How a query becomes an answer
When you type a question into Heinrich, this is what happens.
The bridge receives the text. It calls a parser that resolves words and phrases to concept identifiers, using the label index to find which addresses in the field this question actually refers to. The resolved seeds are passed to the Mind as an IntentPacket.
The Mind seeds those concepts with amplitude 1.0 and begins propagation. Energy moves through three pathways. Same line propagation through canonical bias ratios to neighbouring concepts on the same harmonic line. Cross line propagation through composite signature coupling to lines that share resonant structure. Cross layer propagation through root frequency ratios across the layer stack. Each step decays according to the strength of the harmonic match.
The cascade is bounded. A 25 second wall clock budget. A cap of 50 transmitters per round. A top K of 20 activated concepts at the end. The result is an activation cloud, a structured set of concepts with amplitudes, phases, and a full reasoning trace.
That cloud is the answer in mind form. It is not a sentence. It is a set of coordinates with values.
The bridge passes the cloud to the Language Cortex through a typed contract. The cortex runs a fact extractor, picks an opener from a verified IsA claim, selects supporting facts by intent, groups them by predicate, and composes one paragraph. Every clause in that paragraph traces back to a specific fact in the activation cloud. No clause exists without a source. Unsupported claims are rejected before the sentence is built.
The cortex returns the sentence to the bridge. The bridge returns it to the UI. The UI shows it to you.
Every step of that path is recorded. Every fact that contributed to the answer is named. You can ask the system why it said what it said and get a physical answer. Not a confidence score. A list of concepts, addresses, amplitudes, and the propagation pathway that connected them.
The shape of a Heinrich install
A complete Heinrich install on a developer workstation is, today, roughly 90 to 120 GB on disk, a few hundred megabytes of Python code, and zero external dependencies in the cognitive path. It runs on a CPU. It does not phone home. It does not stream parameters across an interconnect. It does not need an internet connection to answer a question about Belgium.
The first physical product, HAVEN Ear, ships a curated subset of the same substrate at under 50 megabytes. The same software architecture. The same six dimensional address. The same separation between Mind, Language, and Bridge. Different deployment, identical shape.
This is what we mean when we say Heinrich is software. It is not a service we sell access to. It is a system we ship. The mind is on your machine. The knowledge is on your machine. The reasoning is on your machine. You can inspect every part of it.
The intelligence is Heinrich. The voice is VOXIS. The presence is Haven. All three live with you.
Engineered for Presence.
Stay in the loop
EMPHOS publishes twice a week. Product updates, research, and the thinking behind the build.
Explore Haven, HEINRICH Intelligence, The EMPHOS Vision, All Posts
EMPHOS Group, Chilliwack, BC, Canada, info@emphosgroup.com