← Writing

Compile, Don't Retrieve

Dinike Jayamaha
· 12 min read

Notes on the architecture of long-horizon personal memory

Written in response to South Park Commons' Requests for Curiosity, Summer 2026 — specifically the question “What is the future of proprietary intelligence?”, and its three parts: what the right architecture is for long-horizon user memory; how personal memories will be stored, ported and retrieved if users own them; and what that means for privacy, portability rules, deletion rights and the standards that may emerge.

1 · The question

The intelligence layer has stopped being the interesting part. It is roughly the same intelligence everywhere, from a handful of providers, and it gets cheaper every quarter. What does not commoditise is the corpus underneath it — the record of one person's life, which nobody else can have and no provider can replicate.

That corpus is proprietary by definition. Yours is yours. And nobody has solved what it should be: not the schema, not the retrieval, not the maintenance cost, not the ownership. The frontier labs treat memory as a feature bolted onto a chat product — a scratchpad of preferences that decays, cannot be inspected, cannot be exported in any form that would be useful somewhere else, and cannot be trusted with anything consequential.

I have spent the last three months running a different answer against my own life, full time, as its only maintainer and most demanding user. This memo is a field report, not a survey. Where I have got something wrong or not yet built it, I have tried to say so in the same paragraph rather than the footnotes.

2 · Why I have an opinion

I ran businesses for eighteen years before I wrote software. Group treasury at a multi-billion-dollar apparel manufacturer. An automotive components company. Then Ethical Extracts, the botanical extraction company I founded in 2016 and took through a 2019 transaction that brought together MANE — a top-five global flavour house that had never done business in Sri Lanka — and Akbar Brothers, the world's largest Ceylon tea trader. I stayed five years as CEO through the integration.

Every one of those businesses ran on institutional memory that lived in a few people's heads. How an exception got priced. Who to call when a line went down. Which certification actually mattered to which customer. None of it was written anywhere a machine could act on, and when those people left, it left with them.

Then it got personal. I have email going back to January 2007 — nineteen years — and I could not answer a specific question from it. Not “search for a keyword.” Answer a question. The record of my own life was write-only.

So I built the fix for myself, and I have lived inside it since April.

3 · The architecture claim

The industry's default answer to personal memory is retrieval: embed everything, search at question time, stuff the results into context, generate an answer. Pay per question, forever. The answer is assembled at runtime from fragments, which means it is a guess — sometimes a good one — and it cannot cite itself in any way a person can check.

My claim is that this is the wrong architecture for long-horizon memory, and that the right one is compilation: an LLM reads the raw material once and writes durable, interlinked, human-readable pages — a page per person, per company, per project, per topic — with every claim citing the source file it came from. Raw material is immutable. Pages are rewritten only when new material arrives.

Four consequences fall out of this, and all four are load-bearing.

Economics. You pay once per fact, not once per question. Reading a compiled corpus is nearly free. My own instance is now 700+ compiled pages, as of 28 July 2026, drawn from more than 32,000 emails across nineteen years, plus calendar, portfolio, health and meetings — and it runs for cents a day on commodity flash-tier inference.

Currency. A compiled page can be superseded; an embedding cannot. Decay is the failure mode nobody designs for — a memory that was true in 2019 and quietly isn't now. Retrieval has no mechanism to notice: it returns the stale fact and the current one with equal confidence, because it never adjudicated between them. Compilation forces the adjudication at write time, and a contradiction brake queues the conflict rather than silently keeping both. Being wrong is a bug you can fix. Being stale is a structural property of every memory system that can only accumulate.

Inspectability. A compiled corpus is pages a human can open, read, and correct. When the machine is wrong, you can see where and fix it. An embedding index is a black box that can only be re-queried, never audited.

Provenance. Every sentence in every page carries the source it came from. This turned out not to be a compliance feature. It is the product. An uncited memory is a rumour, and nobody will run their finances, their work, or their relationships on a rumour. Provenance is what upgrades “an AI that chats about your stuff” into a system of record.

4 · Three things I did not expect

Maintenance was the abandonment cause all along. I have seventeen years of dead systems behind me — Evernote, Notion, Obsidian, paper — and I abandoned every one inside two months, always for the same reason: upkeep is boring human labour, and the backlog of unfiled life compounds until the tool becomes an accusation. Every note-taking company has treated this as a user-discipline problem. It is not. It is a labour problem, and the labour just became a machine's job. That is the entire “why now,” stated mechanically: long context windows and cheap inference pushed the cost of maintaining a person's knowledge graph below the cost of the person maintaining it, for the first time ever.

Memory must be able to un-remember. A system that can only add is not a memory; it is a landfill. The engine runs autonomous nightly synthesis — finding connections, filling gaps, fixing rot — and every write it makes goes through a reversibility journal. On 14 July it caught links it had written badly the night before and reverted them, unprompted. That was the moment I believed the architecture, because self-correction under provenance is the property none of the retrieval systems can have: they cannot revert what they never explicitly wrote.

The proactive layer is the product. I built this to answer questions. What I actually use is what it surfaces without being asked: the promise with nothing behind it, the person who went quiet, the connection between a meeting last week and an email from 2019 that no one would have drawn. Query answering is table stakes. A memory that thinks about you while you sleep is the category.

5 · Stored, ported, retrieved

Storage and retrieval I have opinions about because I have built them. Portability is the one I have least right to be smug about. Taking them in order:

Stored as plain text. Every compiled page is markdown with a small structured header and inline links — no database, no proprietary container, no index that only my code can read. A person could open the entire corpus in a text editor and understand it. That is not nostalgia; it is the only storage decision that survives the company that made it.

Retrieved by having been written already, which is the argument above.

Ported — and here is the actual answer, which is that the export problem largely dissolves if you got storage right. There is nothing to convert. The storage format is the export format. Most memory products need an export API because their memory lives in a shape only they can read; a corpus of files does not. I think this generalises. Whatever a user-owned-memory standard turns out to be, it will be a format, not an API. An API is controlled by whoever hosts it, and a right you can exercise only through someone else's endpoint is not ownership — it is permission.

The honest part: my own export button is not shipped. It is disclosed on the site, it is an open obligation under GDPR Article 20, and my own launch gate refuses to let me invite a founding circle until it exists. Deletion — the harder half — does work, verifiably, to zero bytes, with confirmation. I raise the gap on my own initiative because a memo about user-owned memory written by someone who has not shipped export should say so before someone finds it.

On the regulatory question: the direction of travel favours this architecture, and not by luck. Portability rights, deletion rights and access rights all assume a corpus that can be produced, read, and destroyed on demand. A compiled, cited, file-based corpus satisfies all three literally. An embedding index satisfies none of them honestly — you cannot show a user what it knows, you cannot prove you deleted the part derived from a deleted source, and you cannot hand it to a competitor in a usable form. Compliance is usually treated as a tax on architecture. Here it is a test that most of the field quietly fails.

6 · The top-down case

This is the section I have been avoiding, because for three months everything I made was bottom-up. South Park Commons' own writing on -1 to 0 says the builder's artifact is the demo and the investor's artifact is the memo, and that skipping the memo is how builders end up at a local maximum. Fair. Here is the memo's memo.

The direct market is a consumer subscription. I am not going to publish a price here — the price is not set, and I would rather test willingness to pay than anchor it in an essay. The shape of the arithmetic is not in doubt, though. This is a category where a great many people already pay ordinary consumer-subscription prices for tools that remember nothing about them between sessions. A business of consequence is a modest share of the people already paying, and the honest wedge is narrower still: email-heavy knowledge workers with a decade or more of digital history, for whom the corpus is already valuable and already drowning them. I do not need the whole market. I need a small share of a population that has already demonstrated it will pay for a tool that forgets it.

What makes a consumer price viable is the cost curve, and that is a direct consequence of compiling rather than retrieving. The point is not only that it gets cheaper as models get cheaper, though it does — its dominant cost is inference it has already finished paying for. The point is that it is predictable. Compile-once makes cost a function of how much life a person brings, which is knowable when they arrive and bounded thereafter. Retrieval makes cost a function of how much they ask, which is knowable only in arrears and bounded by nothing. For a consumer subscription the second is the harder problem, and it is not the average that kills you — it is the variance.

The real constraint on this market is not price and not capability. It is trust, and it took me a while to see that it is the business rather than an obstacle in front of it. This product reads a person's private email. Strangers do not grant that to an unknown company, however good the demo is. That is why the founding circle is ten people and not a hundred: finding ten who will hand over their mail is a genuinely harder test than finding a hundred who would install a toy, and whoever solves the trust problem for personal memory has solved the category, because everyone faces it and no amount of model quality substitutes for it.

The larger prize is not the subscription. If the compiled, cited, user-owned corpus becomes the standard form of a person's memory, it becomes the substrate every assistant — including ones I do not build — has to read from. Storage, portability and retrieval of a life's knowledge is infrastructure, and infrastructure is where a good product becomes a generational company.

And the same engine pointed at a company instead of a person is a second product with a shorter path to revenue: an institutional-memory layer for SMEs — the exact problem I watched four businesses fail to solve — already in early testing with a financial research firm as design partner. I am deliberately building the personal product first. It is the harder trust problem, and solving it makes the company product nearly free.

7 · Why the incumbents don't take this

Google, Apple, OpenAI and Anthropic all have memory features and unbeatable distribution. The honest structural argument for why they still lose this category:

Their memory exists to serve their assistant. Mine exists to serve its owner. Those sound similar and are opposites. A platform's economic gravity pulls toward memory as lock-in — opaque, non-portable, resident inside their product — because portable memory makes assistants interchangeable and destroys switching costs. User-owned, exportable, inspectable memory is not a feature they are late to; it is a feature their business model punishes. The companies best positioned to build the corpus are the ones most disincentivised to let you own it.

There is a second reason, which is silo shape. A platform will compile its own data deeply and a competitor's not at all. A person's life is not shaped like any one vendor's silo, and the memory that matters most tends to sit across the seams.

The honest risk that I am wrong: in consumer, distribution has beaten architecture before, and a platform could ship a good-enough opaque memory that most people accept. My bet is that “good enough” fails specifically at the consequential end — money, work, health, the things you need citations for — and the consequential end is where subscription willingness lives.

8 · Ideas considered and discarded

The financial-analyst product. I built it first: a pipeline compiling regulatory filings for all 283 companies on the Colombo Stock Exchange nightly, with cryptographic provenance on every figure, plus working macro models. It works, it is legible, it is the easier sell. I set it aside as a company because it is a feature of the architecture, not the architecture — and because the personal corpus is the trust problem that, once solved, makes every vertical cheap.

Brain-for-SMEs first. Sequenced second, deliberately, for the reason above.

A better note-taking app. Never. The world does not need another one. The insight is that notes were always the wrong unit; the unit is the compiled page, and no human should be writing it.

Capture hardware. Premature. The corpus that matters already exists in mail, calendars and files; the bottleneck is compilation, not capture.

9 · What would have to be true

I wrote down the numbers that will tell me to stop before I had any users to measure. A ten-person founding circle onboards in September, instrumented from each person's day one. The gates:

If the circle abandons this on the same two-to-six week curve as every note app I have ever quit, then the maintenance thesis is wrong and the architecture does not save it. I would rather learn that in October than in a year, and I would rather have published the thresholds in July than negotiate with myself about them in November.

Conviction is not certainty. Conviction is being glad you made the bet either way. These gates are how I keep myself honest about which one I have.

10 · The local maximum I am most at risk of

A beautiful instrument for people exactly like me — forty-year-old, email-heavy operators who read at night — mistaken for a company. Three months of bottom-up work and one external user is precisely the profile that walks into that trap, and I know it. The test is strangers, and the strangers arrive in September. This memo exists because the top-down muscle is the one I have not trained, and writing it down is how I start.

The system described here runs today at lifewoven.app. The architecture, the safeguards, the two production failure cases and the pre-registered validation protocol are documented in full in the companion technical report, Autonomous Maintenance of a Personal Knowledge Graph.