Institutional Memory, Not Model Access, Is the AI-Native Moat

Garry Tan argues that the gap between modest and extreme AI productivity gains is not explained by access to a better model. People getting twofold gains and people getting hundredfold gains may be using the same Claude model, weights, context window, and API. “The leverage is not in the weights,” he says. “It’s in how you wire the work.”

His own estimate is deliberately provocative. In 2013, while a YC partner building its internal social network, investing, and working nearly full-time as an engineer, Tan says he produced roughly 14 useful logical lines of code a day. He recalls that as broadly normal for the era: perhaps 15 to 50 usable lines a day, not thousands. Today, despite working fewer hours because he has a 5 p.m. child pickup, he calculates his output at roughly 400 times what it was.

Tan invites a severe discount for bloated agent output, scaffolding, and favorable assumptions. Even under what he calls a pathological verbosity penalty, he says the increase remains eightfold at the floor and 80-fold in a middle case. The significance of the number is not raw code volume. It is that the same person can organize far more work when the agent is treated as something closer to a workforce than autocomplete.

YC’s experience, he says, has reinforced that view. A quarter of companies in its Winter 2025 batch had codebases that were 95% AI-generated, and that batch became the fastest-growing and most profitable in YC history, according to Tan. He notes that 94 YC companies have crossed $100 million in revenue after a seed check, but does not claim AI-generated code caused growth. His narrower observation is that the fastest-growing founders are restructuring work around agents.

The durable asset in that restructuring is not model quality, which Tan calls rented. It is a company’s accumulated skills, memory, processes, and ability to select relevant context. In his formulation, the organization that builds a reliable “brain” owns the layer that compounds even as models change.

Garry Tan’s central product thesis is that organizations need a memory layer that keeps them from repeatedly asking what they already know. He calls this a company brain: a library of institutional material plus a librarian that determines what should be brought into an agent’s working context for a particular task.

He starts from a comparison with human working memory. People can generally hold only about seven items in mind at once, he says, invoking the familiar “seven plus or minus two” formulation. Checklists, filing cabinets, org charts, and other institutional tools are prosthetics for that constraint.

An agent with a million-token context, by contrast, can hold roughly a thousand pages. Tan explains the scale to his child as three Harry Potter books open at once: the system can find a needle in any of them and synthesize across them in seconds. Whether that constitutes AGI is beside his argument. It is already a substantially different operating regime from one designed around a human’s limited working memory.

Yet three books are still not a company. A company is a far larger library: emails, meetings, customer conversations, decisions, postmortems, and the reasoning behind them. The important question is who decides which material belongs in the active context. That is context engineering.

Tan describes retrieval as the underlying primitive, conceding that this can sound like “just RAG,” in the same way that Postgres can be described as “just B-trees.” The difficult work surrounds retrieval: deciding what gets written into the knowledge base, how it is enriched and linked, what becomes hot memory rather than cold reference, and how conflicting records are resolved. “Being worth retrieving from,” he says, “is the product.”

His own open-source project, G-Brain, is meant to provide that retrieval layer across different agent harnesses. It selects the relevant “three books” for a given task. Tan says his personal instance has grown to about 220,000 pages, mostly written by agents from his email, meetings, 20 years of notes, and lived experience. When a founder emails him about a crisis, he says the system can retrieve that founder’s earlier correspondence, three portfolio companies that faced a similar problem, and what worked for them before he has finished reading the message.

That is the line he draws between an assistant and a colleague. A colleague, in his description, acts with access to what its human counterpart already knows.

Garry Tan emphasizes that a company brain is not useful simply because it contains a lot of material. Its failure modes are straightforward: an uncurated brain becomes “a garbage dump with great search”; retrieval can surface stale facts with total confidence; and a bad skill file can preserve a bad process indefinitely.

The required primitive is therefore memory plus hygiene. Tan calls for provenance on every fact, contradiction checks when new information collides with old information, and a librarian—human plus agent—whose job includes pruning. The brain should be treated as production infrastructure, not as a passive archive. Otherwise, the organization gets an agent that is wrong in ways no one can trace.

That discipline also applies to daily agent work. Tan’s operational rule is “never do one-off work.” An agent may produce an inadequate first draft, much as an intern might. The user can correct it and ask for another attempt. But the work is not complete once the immediate output is acceptable. It should be converted into a reusable procedure.

Tan calls that conversion “skillify it.” He points users to a skill file that can turn a completed agent task into a reusable skill for their harness. The objective is to capture successful work as an organizational capability rather than repeat the same instruction later.

An organization that does this gets smarter every day, he argues; one that does not wakes up with amnesia, regardless of the quality of its underlying model. This is why the memory and workflow layer matters strategically. A stronger model can be rented by competitors. The accumulated library, the procedures built from prior work, and the system for maintaining them belong to the organization that created them.

Garry Tan maps the components of an agent workflow directly onto the components of an organization. A skill file is an employee: one capability, one job, described clearly enough to execute. A resolver table is an org chart. When a task arrives, the table decides which specialist should handle it and which files should be loaded.

He gives the example of an agent workflow in which a request to alter a test directs the system to load tests.md. Filing rules are the organization’s internal processes. Trigger evaluations function as performance reviews: tests that check whether the intended file was actually loaded and whether the procedure was followed.

The claim is intentionally literal. The functions once assigned to a much larger organization can be represented in markdown files, rules, evaluations, and some conventional code. An AI-native company, in Tan’s account, does not merely add agents to existing sales, support, operations, and finance functions. It encodes those functions as written procedures for agents to execute, then employs engineers to maintain the procedures and handle work the skills cannot yet do.

He says that can extend as far as filing taxes, assuming the relevant procedure has been encoded well enough. The personnel model changes accordingly: the important employees are not only conventional programmers but people who can maintain skills, construct workflows, and supervise agent output.

Tan cites two YC companies to illustrate the resulting revenue-per-head economics. Emergence, an AI app builder from YC’s Summer 2024 batch, went from public launch to nine figures of annual recurring revenue in eight months, he says. When it crossed $15 million in ARR, it had 15 people. Retool, which Tan identifies as Winter 2024, was at $60 million with roughly 40 people.

Tan characterizes that revenue per employee as unlike what had existed in earlier software companies or more capital-intensive sectors. These companies are not, in his telling, exceptional because they hired unusually few people. They were built around a different organizational physics.

YC is attempting a similar shift internally. Tan says people in media, events, and finance who had never opened a terminal are building skill files and cron jobs. One finance employee consolidated roughly 100 Excel workbooks into a single application using YC’s internal Open Claude and company brain. She was not operating as a conventional programmer, he says, but as a manager of agents.

Garry Tan divides computation into latent space—the LLM itself—and deterministic space, where conventional software stores state and executes predictable operations. He argues that many AI engineering problems arise because computation assigned to one space is instead made to happen in the other.

Latent space is for taste, judgment, interpreting vague requests, and other nondeterministic tasks. The model is steered through markdown instructions, but it performs the human-like interpretation. Deterministic space is where code agents write and run TypeScript, Erlang, or other conventional code, and where structured state should live.

Tan’s example is seating participants at Startup School. YC expects to work with 6,000 people and may attempt to seat 800 at a time so each participant is placed beside especially useful people to meet. An LLM can contribute the human judgment: which attendees are likely to benefit from being near one another. But the actual representation of 800 seats in a multidimensional array cannot sensibly reside in the context window. That state must be managed deterministically.

The arrangement combines both spaces. The model does the matching work that resembles human social judgment; deterministic software stores and manipulates the seating configuration. A human attempting the same task might print hundreds of pages and physically arrange them in a room. Tan suggests an agentic system could potentially do it in about 10 minutes with a few hundred dollars of tokens.

Garry Tan’s advice to founders has two parts. The first is to build an AI-native company from the beginning: keep the team thin, retain the founder’s connection to the code, create skill files for recurring work, and begin accumulating a personal and organizational library from the first week.

He recommends G-Brain as a free, open-source option, but explicitly argues that the concepts matter more than his repositories. Claude Code is his preferred “Ferrari”; Codex, he says, is a “really good Honda” that can accomplish 90% of the work. The operating model travels across tools: distinguish latent from deterministic computation, build skills, maintain a library and librarian, and avoid one-off work.

The second opportunity is to build the infrastructure underneath that model. Every company, Tan argues, is about to need a brain: a memory layer, a personal AI that actually knows its user, and a system that can select the few relevant documents from an immense body of history. He says he is building G-Brain openly and does not intend to monetize it because he believes the foundational layer should be open in the way Linux is open. But company brains, personal context, and the librarian function remain open territory for companies to define.

He frames the potential in terms broader than startup efficiency. Tan acknowledges anxiety about job losses, but calls the belief that this necessarily produces a jobs crisis a failure of imagination. In his view, abundance arrives through shipped software that multiplies what people can do.

His closing example is a friend with a son who has a rare form of epilepsy. The father built a repository of 80,000 markdown files—a company brain centered on one child—and used it to push toward the limits of available knowledge about the condition. Tan presents that not as an aside but as the same architecture in a personal setting: a library, a librarian, and the right material opened at the moment it matters.