Personal Models: how we are designing an intelligence that learns one person
Foundation models know the world. They do not know you, and each conversation with one starts close to zero. We think the next hard problem in AI is not a model that knows more. It is one that learns a single person accurately across years, notices when that person has changed, and stays theirs.
We call the program Personal Models. This note is the design, written before the code. Nothing below is shipped. What Anjo does today is at the end.
The bar
We wrote the bar before the architecture, so the architecture could fail it:
After years together, Anjo understands what matters now, notices when you have changed, and can carry that understanding into a new situation. You can correct it, take it with you, and keep it working without us.
And the harder sentence, which is the ambition of the company:
You could give another AI everything Anjo remembers about you, and it still would not know how to be your Anjo.
That is a claim about learned judgment, not lock-in. Your export is meant to be complete. A capable competitor can inherit it. Anjo only keeps its place if it learns you better every month than anyone starting from that export, and we intend to measure that gap in the open.
The person is not the evidence
The first design decision is a sentence we repeat in review: the person is not the evidence. The evidence is an incomplete, permissioned record of what someone chose to share.
So the durable object is not a set of weights. A Personal Model is a versioned bundle of four parts:
- Evidence: source records, corrections, outcomes, provenance, and retention state. You can inspect, correct, suppress, delete, and export it.
- Understanding: facts and uncertain hypotheses about your present context, each with a valid time, supporting and contradicting sources, and a status. "Used to want X", "wants X at work" and "wants Y now" can coexist without being averaged into a permanent trait.
- Continuity: commitments, the history of your conversations with Anjo, and the boundaries you set.
- Learning contract: what may be used, what you have forbidden, and what the current build depends on.
Weights, latent vectors, summaries and prompts are rebuildable representations of that bundle. If a representation is lost, or a model is retired, the bundle is what survives.
What may be learned from which signal
Most personalization systems quietly learn from whatever is easy to count. Our design names the forbidden inferences explicitly:
| Signal | May be used for | May not be read as |
|---|---|---|
| An explicit correction | Applying it now, in its scope | Evidence about your life |
| "That helped" | Satisfaction with a reply | Proof the advice was right |
| Your report on a criterion you set earlier | Labelling that outcome | Proof Anjo caused it |
| Return visits, session length, silence | Product diagnostics | A reward for learning |
The last row is the one that matters. Engagement never becomes a learning objective. Someone who needs Anjo less can be a successful user.
Correction before training
The first milestone contains no training at all. It is four explicit preferences: no unsolicited advice, fewer questions, shorter replies, and "help me think" versus "just listen". Each has a scope: this reply, this conversation, or from now on.
The demonstration is deliberately small. Correct Anjo once, and the next relevant reply changes. Restart the server, open a new conversation, and it still holds. Ask for an exception, and the standing preference survives it. Delete the preference, and it stops applying, even though the original sentence is still in your history.
We start here because a learned policy that cannot beat an explicit, correctable preference has not earned its complexity. Explicit corrections bypass training and take effect immediately, and declining any future learning trial will not make correction stop working.
Earn every parameter
Per-user adapters and hypernetworks are active research directions, and papers show learned per-user representations helping on particular tasks. None of that shows they beat a strong baseline for a changing person with sparse evidence. So the design treats capacity as a budget, not a ladder:
- Explicit state, good retrieval, and a compact prompt. Everyone gets this, and it is the baseline every experiment must beat.
- A small learned policy, only where that baseline misses a stable pattern, with calibrated abstention.
- A per-user adapter, only if it beats the simpler candidate by enough to justify its build, serving, and deletion costs.
A person may stay on explicit state forever. If a learned representation cannot beat the baseline on fresh data, it is cut, and the vision is not.
Deletion has to survive a hostile demo
"Wrong", "changed", "do not use" and "delete" are four different operations. Delete removes the source data, invalidates everything derived from it, stops affected reads, and rebuilds from a clean base without it.
We are also specific about what that does not prove. A rebuild without your data is not a mathematical proof that no output can resemble it, and a copy you already exported is outside our control. We will say those limits in the product rather than promise "erased forever".
Ownership has a test we intend to run: restore an exported Personal Model on a clean machine, block Anjo's domains, and keep talking, correcting and deleting. An export that only works while our servers are up fails.
What would prove us wrong
We committed to disconfirming evidence before the first experiment:
- A strong explicit-state and retrieval baseline matches the learned policy.
- People cannot perceive the improvement in blind comparison.
- Corrections stay burdensome.
- The learning curve flattens with tenure.
- Serving and maintaining it costs more than people will pay.
The first result kills an unnecessary representation. The rest require a product or business change, not a larger training pipeline.
Where this stands
Today, Anjo on iPhone remembers across conversations, keeps a dated receipt behind every read, lets you mark a read wrong, filters invented recollections before display, and scores 0.821 recall@5 on LongMemEval-S. Nothing you say trains Anjo-owned models.
Personal Models is a design and a staged research plan. The first milestone, explicit corrections that hold, is specified and not yet built. Any evidence collection for learning will be a separate, opt-in trial with its own notice, and it will not start before the public wording has been reviewed. Results, including the ones that go against us, will be published here.