Exact, semantic and structural memory — in one field.
Resonant Field Storage holds a document three ways at once: the exact bytes, sealed and tamper-evident; what it means; and who did what to whom, when, where, how and why. One representation, three ways to ask it, and measured zero interference between them. There is no index, so there is nothing to rebuild.
Ask a business where its knowledge lives and you get three answers. A document store holds the exact files. A vector database holds what they mean, approximately. A knowledge graph, if anyone built one, holds the structure. Three systems, three copies of the same thing, and three ways for them to disagree.
This is one field. A document is written into it once, as a single complex vector with three bands, and each band answers a different kind of question. Give me the exact bytes back and prove nobody changed them. Find me documents that mean something like this. Tell me who the agent of this event was.
And it keeps ambiguity alive. “John sees Jane with a telescope” has two readings. An embedding commits to one and destroys the other. The field holds both, returns the right one when you ask, suppresses the wrong one completely — and can still hand you the alternative.
No index, so nothing to rebuild
Storing a document is field = field + document. That is the whole operation.
A vector database keeps an index because it cannot afford to compare your question against everything it holds. So every insert rewires a graph of neighbours or shifts a set of cluster centres, every delete leaves a tombstone, and sooner or later the index is rebuilt or its accuracy drifts.
Superposition is linear. The order documents arrive in never matters, nothing already stored is touched when something new arrives, and removal is a subtraction. There is no structure to maintain because the field is the structure.
The problem
An embedding cannot tell who has the telescope.
Modern retrieval turns text into a point in space and finds nearby points. That works remarkably well for “about the same thing”. It is structurally blind to the difference between two sentences made of the same words.
The telescope is how John sees
It belongs to the verb — the instrument of the seeing.
The telescope is something Jane has
It belongs to the noun — a description of who was seen.
One point for both
The words are identical, so one reading wins by accident and the other is gone. Nothing records that there was a choice.
Now put that sentence in a contract, a clinical note or an incident report, and ask a governed system to act on it.
Three systems where there should be one
Nothing makes them agree.
The document store
Holds the exact bytes and knows nothing about what they mean.
The vector database
Holds approximate meaning, loses structure, and re-indexes as it grows.
The knowledge graph
Holds structure, when somebody builds and maintains it, separately from both.
One field holding all three is not an optimisation. It removes the question of whether they agree.
How it works
Three bands, one vector, zero crosstalk.
A document becomes one complex vector of 3,072 dimensions — three bands of 1,024 — placed in disjoint slices with guard bins between them.
Exact
The document itself, sealed with authenticated encryption. Ask for it and you get the bytes back exactly, with any tampering detected. A hundred round trips out of a hundred under load, and 3.1 times the capacity margin the requirement asks for.
Semantic
An embedding carried as a waveform. The band that answers “find me things that mean something like this”.
Structural
The event frame: who did what to whom, when, where, how, why and by what means — sixteen semantic roles, each bound to its filler by complex multiplication and summed into one vector.
Why the bands do not interfere
Measured: energy in the guard band, 0.0. Leakage between bands, 0.0.
The separation gates record zero leakage violations, zero overlap violations and a minimum guard thickness of eight bins. That is what makes one field safe to query three ways — a semantic search cannot corrupt the exact bytes, and a structural query cannot pick up noise from the meaning band.
To ask “who was the agent”, multiply the frame by the conjugate of the agent vector and clean up against the known fillers. That is exact algebra rather than a similarity guess. The algebra itself is published work — we cite it rather than claim it. What is ours is carrying it in a band of a field that also holds the exact bytes and the meaning, with the separation measured at zero.
Meaning in superposition
Both readings held. The right one returned. The wrong one suppressed.
This is the part no vector store can do, because it depends on structure surviving superposition.
A binding's carrier is the predicate's phase multiplied by the role's phase, so “telescope as the instrument of seeing” and “telescope attached to Jane” land in different places in the field. Both are added into the same vector, and because their carriers point in nearly orthogonal directions they coexist without corrupting each other.
Every role vector is generated from a hash of its own name and tested against every role already placed; if its overlap with any of them is above 0.14 it is regenerated. Near-orthogonality is constructed, not hoped for — and anyone can rebuild the same codebook from the names alone.
Seven hundred and eighty evaluated pairs, all four gates passed, no catastrophic collapse under superposition.
And it knows may from must from must not
For anything that decides on an action, that distinction is the entire question.
Modality is multiplied over the whole frame: asserted, possible, obligated, question, imperative, negated. “The system may delete the record”, “the system must delete the record” and “the system must not delete the record” are one proposition in three modalities.
An embedding places them close together, because they share nearly every word. The field separates them, because the modality is a factor of the whole meaning rather than one more word in it.
Evidence
Structure beats embeddings. And the field knows its own limit.
Measured in the repository.
A capacity limit it states, proves and refuses
It refuses the exact point the theory predicts it breaks.
The limit is an inequality over the parameters you control, checked by a proof kernel in exact whole numbers with every rounding set against the claim. It certifies two configurations, proves two refusals — and refuses 13,700 documents, the predicted breaking point, because at that point the margin is zero.
Measured against it: accuracy holds where it certifies and degrades where it refuses, never before the predicted point. The bound under-promises, which is the direction a safety limit should be wrong in.
Predicted before it was measured
On a corpus with relevance judgements we did not write.
Hard questions
What a search or data platform engineer asks first.
Answered as we would answer them in the room.
“Why not just use a vector database?”
For approximate similarity at scale, use one — they are mature and operationally excellent.
What none of them does: hold the exact bytes with tamper detection in the same representation, answer a structural question with algebra rather than a similarity guess, keep both readings of an ambiguous sentence, separate may from must not, or grow without an index to maintain.
“Isn't role binding in vectors an old idea?”
Yes, and we cite it. The binding algebra is published work and we do not claim to have invented it.
What is new is three bands in one field with measured zero crosstalk, roles made near-orthogonal by construction, modality factored over the whole meaning, and ambiguity held in superposition with the results measured.
“Why does a governed system need this?”
Because a decision is only as good as the evidence it was made on. If the memory under an agent approximates, flattens an ambiguous clause into one reading, or quietly shifts as it re-indexes, every control above it is enforcing a rule against the wrong facts — and the record will look perfect.
In the fabric
The encoder decides what a document means. The field holds it.
The Nota Memoria Engine turns text into frames: a predicate, up to twenty-two semantic roles, a modality, and a bounded affect gain. Every codeword is derived from its own name, so the encoder and the field agree on the geometry without shipping weights between them.
What crosses from encoder to field is validated and refused on any mismatch — dimension, magnitude to within one part in a million million, encoding version, duplicate source identifier.
The thesis
Mathematics as the nervous system of AI.
Everything here descends from one argument: that the integrating substrate for artificial intelligence should be mathematics itself — not another orchestration layer, not a better prompt, and not a policy document.
Each part of a modern AI system works. The joins between them do not. Vision, language, planning and retrieval are each remarkable and they are integrated through hand-built pipelines and brute-force scaling. The thesis proposes a shared mathematical space that components write into and read from through operations defined once and behaving the same way for all of them — a nervous system rather than a bundle of wires.
Guarantees become measurable. Every property claimed has a quantity attached. Measure it and either the implementation holds or it is broken; there is no third answer.
The foundation is reusable across customers. The calculus, the construction engine, the control plane and the receipts are common. Your rules, connectors, integrations and authority model are yours.
The ladder, in order
Each rung was built from the one before it.
That order is why the components share a foundation instead of being a suite assembled after the fact, and it is why a refusal at the action boundary can be traced back through a proof to a sentence somebody wrote.
The paper
Openly licensed, so you can check the argument.
Open
Mathematics as the Nervous System of AI: A Unified Field Operator Framework for Distributed Cognition. Philip Siniscalchi, v9, 27 August 2026, CC-BY-4.0.
Falsifiable
It separates conformance — does the implementation obey the mathematics it claims — from superiority over alternatives, and refuses to let the first stand in for the second.
Bounded
No claims about consciousness or sentience. The biological analogies are engineering inspiration, not identity claims. Theoretical extensions are labelled as a roadmap, never as capability.
Start at one seam
Twenty minutes, on your own machine.
Connect your own coding agent, ask it to do something that writes, deny it, verify in your own environment that nothing happened, then read the ledger.