Write a fact once. Get the same fact back.
Write it once. Get the same fact back — the original, not a cousin.
Write it once. Get the same fact back — the original, not a cousin.
A filing system, not a neighbor
Trinity Sky is a store a company owns. A fact is filed under a key. Later the same key is asked for, and the same object comes back — the options, the price, the why — not a sentence a model judged probable enough to keep the conversation moving. Retrieval-augmented generation and a sliding window rank nearby text. This archive performs an algebraic get: it unbinds the value bound to the key. The rest of the stack may still write prose. The store is the product. When the store returns the object that was filed, the door a person opens and the sentence a model writes both sit on a record the company can keep.
That is the want, stated once. A strategy call, a pricing change, a clause in a contract — filed once, asked for later, returned as the same object. Last quarter’s decision remains the sentence a committee can open. Age leaves a fact where it was filed. This morning’s work sits beside March. Both are reachable. The address is the binding.
The product is not a longer window. It is a cabinet that recovers the object it stored. A general counsel who opens a clause should receive that clause. A pricing committee that recorded a why should receive that why. Fluency can still sit on top. Fluency is not the cabinet. A partner can forward this page because the claim is a single sentence a board already knows how to use: file it, ask for it, receive it.
Bind once. Unbind the same object.
Associations live as structured patterns in a Fourier holographic reduced representation — FHRR. Binding is composition: a key and a value become one object in a frequency-domain archive. Recall is the inverse of that composition. The operation that recovers the value is exact unbinding, not a similarity ranking. There is no token replay of the archive. The buyer is not charged by the token to remember who they are.
For a single stored association, unbinding recovers the item that was written. When many facts share a trace, quality depends on how full the store is, not on where a fact sat in a prompt. Every stored item is equally reachable. There is no middle to lose, because position in a prompt is not an address. ChatGPT, Claude, Gemini, Copilot, Cursor, and every RAG stack can ask what is close to a query. They cannot ask what was bound to this key and receive the filed object. Their indexes support an inner product. They do not support an inverse. Trinity Sky does.
- Bind. A key and a value become one object.
- Unbind. The inverse recovers that object.
- No replay. Reading the archive leaves the past off the token meter.
Bind once. Unbind the same object. If the match is weak, return not found.
The category break is the inverse. An inner product finds a neighbor. An inverse recovers a binding. The first is search. The second is a filing system. A committee that already pays for search is not being offered a longer neighbor list. It is being offered the object it filed.
The right to hear not found
The second half of the product is the refusal. A fail-closed threshold of ϑ = 0.70 sits on the recall path. Below that confidence the archive returns not found rather than a fabricated value. The threshold is a gate, not a system prompt.
The buyer is buying the right to hear not found. A thin trace, a crowded store, a key that matches nothing on file — those are times the archive stays quiet. Silence keeps the record clean. A confident invention would spend the trust the lookup is there to earn.
A boardroom can act on what it filed. It can act on an empty reply. It can act on a fluent stand-in only after a person has already spent the hour checking it. The return is that record, or it is empty. Both answers are the product. Fluency is not honesty. A model that must keep talking will mint a cousin. An archive that may stay quiet will not. The capital question is which of those two machines a company wants in the room when the number, the clause, or the reason has to be the one it filed.
Hardware they already own
The archive lives on the buyer’s hardware — in the building or in the buyer’s own cloud. When the relationship ends, the memory stays. Every write is anchored so that a later change is visible: tamper-evident storage. A later change leaves a mark.
A contract that ends leaves last year’s bindings in the room. A model upgrade leaves the address of a fact where it was filed. There is no token tax to reread who they already are. Reading the company’s own past is a lookup, on machines they control.
That is the difference between renting a window and owning a cabinet. The language model on top can still write prose. The model is not the filing system. A chief information officer can defend the install because the company owns the record: the bytes, a write that stays visible if someone later changes it, and a door that stays shut when the archive is unsure. Sovereignty here is a place the bytes live after the vendor leaves.
The past stays off the token meter, on machines the buyer already owns.
What the book already printed
The book prints a side-by-side of the primitive. NEOMORPHIC SSI figures are from internal validated testing. Competitor figures are drawn from published benchmarks and pricing pages.
| Metric | NEOMORPHIC SSI | GPT-4o | Claude 3.5 | Gemini 2.5 | Grok 3 |
|---|---|---|---|---|---|
| Recall accuracy | 100% | 94% | 92% | 91% | 89% |
| Hallucination rate | 0% | 5.6% | 8.2% | 9.8% | 13.6% |
| Cost per recall | $0.00000044 | $0.20–$0.50 | $0.15–$0.30 | $0.10–$0.25 | — |
Those dollar figures are unit costs from that labeled comparison. Isolated recall on a quiet machine is 13.834 µs; under load we have also seen 5,127 µs against a 200-microsecond budget — one pair, two experiments, shown together. Perfect memory recall accuracy is the primitive: write once, exact get, fail-closed if unsure, no token tax on the past, data on the buyer’s hardware.
What a partner is being asked
Put the pieces together and the company is no longer an assistant with a better prompt. It is owned memory: bind once, unbind the same object, hear not found when the archive is unsure, reread the past without a token tax, keep the bytes on the floor they already own.
$20 million is an ask to put the first of those archives on a buyer’s floor. There are no customer logos yet. Forecasts are a plan. The table is the book’s labeled comparison of the primitive.
Sources: Trinity Sky, Investor business plan, §2 “The core innovation — perfect memory” (docs/gtms/09-raise/BUSINESS-PLAN-INVESTOR.md, August 2026). The inverse, the fail-closed gate, the internal-testing versus published-competitor table, and the ask are restated from that section.