Three days ago Heinrich had 128 concepts. Today it has 1.75 million. The CPU usage is still 0.2%. The RAM is still 78 megabytes. No GPU. No server. No degradation in response time.
This is not an optimization story. We did not find a clever way to compress the data or cache the queries. The efficiency numbers did not hold because we worked hard to keep them — they held because the architecture makes it structurally impossible for them to get worse.
That distinction matters. It is the whole point.
What happened this week
On April 14 we completed the full ingestion of ConceptNet — 1,002,949 knowledge nodes, 282,973 edges, representing one of the most comprehensive structured knowledge graphs ever assembled. Every concept connected to every other concept it relates to, through typed relationships that carry meaning in their physics.
On April 15 we started the Wikidata full dump ingestion. Wikidata is orders of magnitude larger — a structured representation of human knowledge covering every entity, relationship, and fact that Wikipedia's global community of editors has verified and organized. The ingestion pipeline runs at approximately 1,500 nodes per second, six workers in parallel, reading compressed triples from a 90GB dump file and writing them into Heinrich's frequency field.
By April 16 the field had crossed 1.75 million nodes. It is still growing.
At no point during this growth did we measure any increase in CPU usage or RAM consumption during queries. The system that answered "what is a mammal" in 2.5 milliseconds with 128 concepts answers the same question in 2.5 milliseconds with 1.75 million concepts. The field grew by a factor of 13,000. The query cost did not move.
Why this is architecturally inevitable
Every existing AI system — large language models, vector databases, embedding search — has a resource profile that grows with the amount of knowledge it contains. More parameters means more computation per query. More vectors means more distance calculations. More data means more memory. This is not a flaw in how these systems are engineered. It is a consequence of how they store knowledge.
Heinrich stores knowledge differently. Every concept lives at a specific frequency coordinate in a layered signal field. Retrieving a concept is a Goertzel correlation — a single-frequency signal processing operation that takes microseconds of arithmetic on any CPU. The cost of that operation does not depend on how many other concepts are in the field. It depends on the number of concepts that activate in response to the query — the subfield that resonates.
When you ask "what is a dog," Heinrich does not search 1.75 million nodes. It activates the subfield around the dog frequency coordinate and propagates from there. The rest of the field is silent. The computation is proportional to what is relevant — not to what exists.
This is why the efficiency advantage does not erode at scale. At 50 million nodes the query cost will be the same. At 300 million nodes it will be the same. The architecture does not work any other way.
The ingestion pipeline
Building 1.75 million nodes is not a trivial engineering problem. The pipeline that does it runs two passes over the data. Pass 1 creates every entity — assigning each concept its unique harmonic frequency coordinate within its domain layer, ensuring no two concepts share the same address in the same space. Pass 2 wires the edges — reading every relationship triple from the dump and connecting the concepts it links, using the harmonic ratios that encode relationship type in the physics of the coordinate system.
At peak throughput the pipeline processes 227 million triples per run, writing new nodes at over 1,500 per second while the RTX 4060 handles coordinate assignment in the background. The system recovers cleanly from interruptions — emergency checkpoints fire on shutdown, and every restart picks up exactly where the previous run ended.
The target is 50 million nodes. At current throughput, we are less than a month away.
What 1.75 million nodes means for Heinrich
More nodes means more questions Heinrich can answer with confidence. More edges means more causal chains it can follow, more relationships it can report, more derived connections it can surface from the geometry of the field alone.
When we asked "what causes disease" two days ago, Heinrich reported an honest gap — the causal connection was not yet in the field. At 1.75 million nodes, the Wikidata edges that link pathogens, bacteria, viruses, and immune response to disease are beginning to land. The answer is getting built in real time, by the ingestion pipeline, without any programmer deciding what to teach the system.
Heinrich does not need to be told what connects to what. It ingests structured knowledge and the relationships are there, encoded in the harmonic ratios between frequency coordinates. The physics carries the meaning. The field thinks.
What this is not
This is not a database with a fast index. A fast index still searches. Heinrich does not search — it resonates. The difference is not semantic. A search scans candidates and ranks them. A resonance response activates what is harmonically present and returns what the field contains. There is no ranking because there is no scanning. The answer is either in the field or it is not, and Heinrich tells you which.
This is also not a vector embedding system. Vector search computes distances between high-dimensional representations and returns approximate nearest neighbours. The results are probabilistic. Heinrich's retrieval is deterministic — the same query returns the same result every time, because the field is a physical structure, not a statistical one.
What comes next
The ingestion continues. The target is 50 million nodes — the scale at which we will run the first formal accuracy measurements, document the reasoning chain quality, and produce the paper that describes what Heinrich actually is and what it can do.
The efficiency numbers will be in that paper. Measured at 128 nodes. Measured at 1.75 million. Measured at 50 million. The same every time. That is the claim. That is what we are building the proof for.
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