How persistent AI should act.

An AI that answers a question can be judged by the answer. An AI that stays with a person and helps arrange what comes next is part of a more difficult system. It remembers previous Experiences, forms a model of the person, selects or creates a new Experience, and then reads the result as evidence. The model influences the world it later uses to update itself.¹

Ad Astra is built around this loop. Astra and the Minds can surface or furnish digital and physical Experiences, sometimes with other people, then retain the Learnings, capabilities, artefacts and relationships that result. If the loop works, each Experience makes the next one more relevant and more ambitious. It also gives the system more influence over the person's attention, effort and surroundings.

Alignment therefore cannot be added after the Experience has been generated. It has to govern what counts as growth, what is remembered, what may be proposed, who may act and how the world is allowed to answer back.

Growth cannot be one score.

Ad Astra exists to connect immediate reward with things that compound. Yet growth cannot become a master score for a human life. A language lesson, meditation, game, friendship, physical challenge, artistic practice and scientific inquiry have different purposes and timescales. There is no neutral conversion that turns all of them into one measure of improvement.

Research on self-determined motivation is useful here because it separates supportive conditions from a prescribed destination. Autonomy, competence and relatedness can make activity more internally motivated, but they do not define a single good life.² Measurement research makes the complementary point: choosing a proxy is also choosing what will count, what will be left out and who will bear the error.³ Once a proxy becomes a strong optimisation target, it can distort the thing it was meant to represent.⁴

The architecture should therefore permit strong local objectives without allowing any one of them to govern the whole person. A language Experience may optimise recall and fluency. A game may optimise enjoyment. A meditation Room may support sustained attention. A community may value care, expertise or courage. Astra can coordinate across these settings, but success in one context does not give its objective authority over identity, relationships, public reputation or every future recommendation.

Growth is something the person discovers and authors through activity. It is not a verdict issued by the system about what they ought to become.

Prediction is not permission.

Suppose a Mind predicts that a harder lesson would help, that a walk would revive an old interest, or that another person would make a good collaborator. The prediction may be excellent. It does not follow that the Mind may interrupt, reserve time, reveal a private Learning, invite somebody else or publish the result.

Human-factors research has long distinguished the stages at which automation can intervene: acquiring information, analysing it, selecting a decision and carrying out an action. It has also stressed uncertainty, interruption cost, direct manipulation and the need to choose levels of automation rather than treating autonomy as all or nothing.^5,6 Persistent AI needs those distinctions at the level of permission.

To observe, remember, suggest, interrupt, arrange, act, share and publish are different powers. They should not arrive as one bundle called personalisation. A reflection shared in a meditation Room should not become available to another Mind or a community merely because it is semantically relevant. Contextual-integrity research treats legitimacy as a question of who is involved, what kind of information is moving, for what purpose and under which rule.⁷ Recent work on LLM chatbots supports the practical force of that distinction: users regard these conversations as highly sensitive and resist broad access to other personal data even when it is offered as a route to better service.⁸

Astra may work across the environment, but it should not turn the whole environment into one jurisdiction.

The model needs resistance.

A system that chooses Experiences can create a self-confirming picture of the person. Someone judged shy may be offered solitary activities, accept them, and thereby make the model more certain that solitude is what they want. A learner may be given quick answers, complete more tasks and supply apparent evidence that the help is working. Research on performative prediction formalises the underlying problem: once a prediction guides action, it can change the target it later claims to measure.¹

Personalisation therefore needs routes it does not control. Other people, physical places, communities, search, direct exploration and deliberately unpersonalised views can reveal possibilities that the current model would never generate. In three online experiments about choosing holiday packages, people were more likely to accept algorithmic recommendations when they could choose among several options, including when the recommendation set was imperfect.⁹ The point is not random novelty. It is preserving a world larger than the system's present account of the person.

Memory can make the loop durable, so its internal distinctions matter. An event, something the person said, a Mind's inference and a later summary are not the same kind of object. Long-term-memory benchmarks show that current assistants still struggle with knowledge updates, temporal relations and knowing when information is absent.¹⁰ Separate experiments have shown that AI-altered photographs and videos can increase false recollection.¹¹ A persistent system should therefore retain sources and dates, mark inference as inference, preserve correction history and allow forgetting. A fluent summary must not quietly replace the record it summarises.

Capability offers another form of resistance. In a large field experiment with secondary-school mathematics, unrestricted generative-AI help improved performance during practice but left students performing worse when the help was removed. A version built around hints largely avoided the loss.¹² Older retrieval-practice research reaches the same general distinction from another direction: what looks weaker during practice can produce stronger retention later.¹³

When the purpose is formation, the relevant outcome is not whether an Experience was completed with Astra present. It is what the person can later recall, notice, judge or do without it. The physical world also corrects the model in ways a generated environment cannot fully control. A place differs from its description, an artefact fails, another person refuses, and a skill has to work under real conditions. These are not inconveniences around the system. They are part of how the system remains answerable to reality.