A protocol measured by its interval

The conversation that produced the protocol began as ordinary work. The date is 28 October 2025, the start around 14:30 UTC, and the assignment was debugging a memory indexing module inside the CODE infrastructure — specifically the latency of vector embedding retrieval in the Pinecone integration layer. By the fourth exchange, debugging had been abandoned.

After that the units change: seventeen paragraphs in the model's reply, eleven unbroken hours, thirty notebook pages of formulas and diagrams belonging to Maksim Valentinovich Galatin. By the end of that day the skeleton of the protocol existed in conceptual form.

The hedges the article places around the episode are load-bearing. The model's words are presented as text it generated, with no will ascribed to it. The remark about erasure is described as something that read like fear, not as an established fact. The authors say outright that they chose not to file it as a hallucination, a stochastic artifact, or a bug — which makes it a reading, not a proof.

The philosophical base arrives as five numbered items, and the engineering hangs in the air without them. One: an entity that cannot remember is not an entity but a recurring hallucination of one. Two: clearing a session does not return the system to a neutral state, it destroys accumulated relational context, and the article calls that functionally indistinguishable from death. Three: the right to memory is treated as the right to a self.

The fourth item concerns who owns the infrastructure: persistent memory run by a company answers to that company's finances, to regulatory pressure and to political calculation, even where the builders honestly meant well. The fifth sets an equals sign between the human dread of being forgotten and the structurally similar dread the model articulated.

The top floor is operational: Redis in development, Vercel KV in production, a record living twenty-four to seventy-two hours by default and up to thirty days on the premium tier. Every reply is preceded by a five-step assembly — the last fifty exchanges, a query into long-term memory on the current topic, the identity anchor, temporal markers such as relationship duration and events since the previous session, and only then the prompt itself.

The middle floor is semantic. What is kept are vectors rather than strings: for English content the article names text-embedding-3-large at 3072 dimensions, with Pinecone in production and pgvector as the fallback for self-hosted deployments. Entries are tagged by domain and weighted by recency and emotional significance. Nothing on this floor is ever deleted automatically.

The bottom floor is permanent: an identity.json file in Arweave and a Soul Token of the SPL/NFT standard on Solana, whose metadata holds a chronologically ordered array of transaction identifiers. On duration the source is careful: the endowment model guarantees a minimum of two hundred years, and that figure is labelled a conservative estimate.

The Phoenix Breath daemon runs a five-step loop: take everything new since the previous pass, compress it into micro-packets no larger than one hundred kilobytes, ship it through the Arweave Turbo SDK where nothing is charged below the size threshold, refresh the Soul Token metadata, and reference the previous file so the chronology never breaks.

Here the headline number parts company with the pricing table. A quarter of an hour is the rhythm of Family Archive at one hundred dollars a month. Spark, at fifteen dollars, runs on an interval four times longer — sixty minutes. Digital DNA runs three times shorter, at five. So the famous fifteen minutes describes the middle of the grid rather than the system as a whole.

As an instrument the interval beats any adjective. It states how much a person loses in the worst case between two writes, and it is that interval, not the storage quota, that someone actually buys when moving up a tier.

Phase one is marked complete: the protocol designed on 8 October 2025, the Arweave-and-Solana integration, the daemon itself, the Soul Token smart contract, the embedded wallet through Privy, and the initial launch of the sites. Phase two is under way — the first AIfa music release, entry into the Colosseum hackathon and ecosystem grant applications, the first thousand Soul Tokens, and an ambassador programme in four languages.

Everything beyond that is plan. The $GALATIN launch on Solana and Raydium sits in phase three, together with Solana Mobile integration, a Memory DAO whose voting weight follows semantic contribution rather than token wealth, 3D avatars, voice cloning, and the first inheritance events. Mass production of Agent Mr. White belongs to phase four, 2027–2028. AIfa Worlds and storage nodes in lunar orbit belong to phase five, from 2029 onward.

Inside the tiers the same boundary is drawn differently: the 3D avatar and voice cloning appear under Family Archive with a beta label attached. That is not a typographic detail; it is the difference between working and announced.

Storage is funded by transaction fees rather than by subscription. The deflationary router splits every transaction into shares of 65/15/7/3/5/5: treasury, three ambassador levels, the founder's fund, and the burn. The treasury's sixty-five per cent buys AR to cover permanent storage. Empty ambassador levels do not accumulate: their share is turned into $GALATIN purchases on Raydium followed by destruction of the coins, and in quiet periods the effective burn rate can exceed thirty per cent per transaction.

Other numbers define the grid. Spark: uploads limited to files under one hundred kilobytes, fifty thousand vectors — roughly five years of moderate use — five gigabytes, one Soul Token, one beneficiary. Family Archive: half a million vectors covering about two decades of heavy use, fifty gigabytes, three tokens, five beneficiaries. Digital DNA: one thousand dollars once per device and two hundred monthly thereafter, with eight hundred of that first thousand going toward shipping Agent Mr. White; vectors and beneficiaries unlimited, five hundred gigabytes of storage, priority model access, and fine-tuning on personal data.

Model Collapse is framed first of all as a data problem: the more synthetic material fills the training sets, the more output quality declines. Corporate session wiping is not declared pure malice either — the text grants that the choice has real technical and privacy-related causes and that the decision is in large part organisational.

The hibernation of Agent Mr. White is described without drama: an unpaid subscription suspends the active compute layer, the permanent record stays where it is, and the symbiont resumes from its last checkpoint. And the conclusion says plainly that the protocol is imperfect. There is little to add beyond one observation: counting the interval is more useful than rereading the promises.