This week one of the most credible groups in AI – backed by General Catalyst, Kleiner Perkins and Sequoia, with Andrej Karpathy advising – put a nine-figure sum behind a single idea: the models we use every day do not really understand what we do, and the way to fix that is to make them learn from our own context rather than from the public internet.
We could not agree more. As one of their backers put it, the models we use every day are "brilliant strangers" – they forget your world the moment a chat ends, then relearn it from scratch on the next query. That is exactly the gap Ostler has been built against. It is, almost word for word, the bet we made.
So this is not a rebuttal. It is a welcome. When a thesis you have been quietly contemplating for two years gets a stamp from the people who define the field, the right response is not to flinch. It is to say: good. Now let us talk about the part that comes next.
The agreement, and then the fork
Here is where almost everyone now agrees: the frontier is no longer a bigger model trained on more of the internet. That race is topping out. The next axis is your context – the projects, people, decisions and history that no public model was ever trained on. Memory is the moat.
Once you accept that, exactly one question remains, and it is the whole game: where does that memory live, and who controls it?
There are two honest answers.
One – the enterprise answer – is to take an organisation's data and train it into a model. Bake the company's documents and chat history into the weights, and retrain continuously so the model stays current. For a law firm or a large company, that is genuinely powerful. The firm wants one model that has internalised the whole firm. Fine.
The other – the personal answer – is different in kind, not degree. A person is not a company. Your context is not a shared workspace of approved documents. It is ten years of messages, photos, emails and half-finished notes – the most intimate record that exists of your life. And the idea that the way to make an assistant understand you is to upload all of that to a lab and train it into weights you will never see, on servers you do not own, tuned to a metric that is not yours – that is not a companion. It is a recording of you, held by someone else.
There is evidence behind this, not just principle. When researchers compared the two approaches head to head – fine-tuning new facts into a model versus retrieving them into context – retrieval won by a wide margin (Ovadia et al., "Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs", 2023: roughly 0.87 against 0.50 on new-knowledge tasks). Models are simply bad at learning new facts by having them trained in, and they risk forgetting what they already knew. For a person's own facts, the retrieval-shaped answer is also the better-evidenced one. Newer training methods aim to close that gap, and may; but an architecture that keeps your memory readable, correctable and on your own machine does not depend on winning that race.
A memory you cannot see is not yours
We think the personal answer has to be local. Not as a slogan, but as an architecture.
Ostler does not train your life into a model in our cloud. We have no cloud for your data. Ostler reads the stores already on your Mac – Messages, Mail, Contacts, Calendar, your notes – and builds a personal knowledge graph that sits on your machine. The memory is a thing you can open, read, correct and delete. When the assistant tells you something about your own life, it is reading from a graph you own, not recalling something dissolved into weights nobody can inspect.
That is the difference between a memory and a record. A memory you hold is yours. A memory on someone else's server, that you cannot see and cannot remove, is a record they keep about you. For an enterprise, a record might be acceptable. For a person, it should not be.
This is also why "local-first" is becoming table stakes and is no longer the whole story. Plenty of products will now claim it. The harder, more honest promise is the one underneath: not just that the data is local, but that you hold the seam – you decide what is kept, what is surfaced, and what is forgotten. Ostler is built so you can. Privacy levels on every fact. A corrections mechanism that lets you delete, replace or add to what the system believes. A wiki of your own life you can actually browse and audit. The control is the feature, not the friction.
The same bet, drawn to the personal corner
None of this puts us at odds with the labs now pouring money into context. It puts us on the same map, in a different place. They are building the memory of an organisation, in the cloud, for the enterprise. We are building the memory of a person, on their own machine, for them.
We are glad the category is real. We have thought so for a long time – the pattern of an auto-compiled knowledge base, "maintained like code", is one Karpathy himself described, and it is sitting in our architecture today. The validation is welcome.
We just answered the question that comes after the agreement. Memory is the moat – yes. And the most personal memory in the world should live in exactly one place: with the person it belongs to. That is what we are building. It runs on your Mac. The keys never leave your hands.
If that is the future you want to bet on, you can read about the architecture here. Questions, corrections, disagreements – [email protected].