← The Last AI That Knows You

The Last AI That Knows You / Volume IV

The Intelligence That Grows With You

How personal AI can improve decisions through memory and adaptation

Founder Reflection by Kam Oomarow, Lumina Horizon Inc.

2026-11-04 · 8 min read

An Audio Edition is not available for this Reflection.

The Intelligence That Grows With You

The first three Reflections examined representation, the limits of a personal model and the authority an agent receives. They established conditions for trust. A further question remains: what would make a personal intelligence worth trusting? My answer is earned usefulness: better decisions, fewer repeated misunderstandings and progress toward goals the person chooses, after counting the effort of teaching and supervising the system. Control makes such a relationship legitimate. Demonstrable benefit gives someone a reason to continue it. Volume II previewed this instalment as “The Persona Is Not the Person.” That distinction remains its foundation. The title now follows the next question: how can an incomplete model become more useful without mistaking its account of someone for the person themselves? The examples below are hypothetical design tests, not customer evidence or claims about deployed Lumina Horizon capabilities. The cited research supports particular mechanisms and findings; it does not establish the complete experience proposed here.

1 The unit of personalization is the decision

Consider a professional preparing for a week away: client meetings, evenings with friends and a long journey home. A wardrobe assistant could begin with sizes, colours, brands and purchases. The more useful question is what the clothes need to make possible during this particular week. The person wants to look assured, travel lightly and move comfortably between occasions. An admired jacket may travel poorly. Frequently worn shoes may reflect a demanding commute rather than a favourite style. A familiar outfit may reassure them at one meeting and feel too conservative at another. Relevant knowledge connects these details. What is the person trying to achieve? Which conditions are fixed? Which trade-off would they consider worthwhile? An assistant could explain why similar choices serve different purposes, recognize when an old compromise no longer applies and identify the unresolved question that would change its recommendation. Deep personalization would mean assistance shaped by goals, circumstances and decision-making, with enough continuity to improve across encounters. This makes the decision a useful unit for builders. A profile can say that someone likes a brand. A decision record can show why they chose one item over another, what constrained the choice and whether the outcome worked. Those details give the next recommendation something specific to improve upon.

2 Learning needs evidence beyond the choice

A person can state what matters, choose among options, correct an interpretation and report what happened later. These supply different evidence. Choosing a jacket establishes that it was selected. Explaining that its lighter construction makes travel easier supplies a reason. Returning it because the shoulders restricted movement adds a constraint. The selection alone cannot establish lasting usefulness. Continuity can come from conversation context, retrieved records or separately maintained preferences. Every interaction need not retrain the underlying model. Lewis and colleagues demonstrated retrieval-augmented generation for knowledge-intensive tasks, combining a model with retrieved documents. That provides a mechanism for supplying information, not proof that a system understands a personal history. [1] The product test is what changes next. A correction should alter the next relevant recommendation. An explanation of why the mistake happened is insufficient if the person must repair it again.

3 Understanding is demonstrated through transfer

In a personal product, understanding can have a practical test: can the system use what it learned appropriately in a new situation? Suppose our traveller wants fewer decisions during demanding weeks. That could change the number of outfits proposed, the presentation of alternatives and the timing of questions. A coordinated selection might help more than an extensive catalogue. During a quieter week, the person may enjoy exploration. The preference has a condition. Repeatedly choosing simple outfits could reflect an aesthetic, limited preparation time or workplace conventions. A useful system would retain competing interpretations with their supporting context, then ask a targeted question when the distinction matters. Enjoying elaborate choices on holiday would weaken a claim of universally minimalist taste. This account is testable. Does a lighter schedule change the recommendation? Does a longer journey change which constraint matters? Can the assistant explain the relevant interpretation well enough for the person to correct it? Transfer also has boundaries. Reducing wardrobe decisions may help with packing. Carrying that preference into a career decision could hide alternatives that deserve careful examination. A convincing explanation cannot establish that the underlying inference is appropriate. LongMemEval separately tests information extraction, reasoning across sessions and time, knowledge updates, and abstention when an answer is unavailable. These are benchmark tasks, not evidence of understanding a life. They reinforce a useful design distinction: sustained assistance requires more than recall. [2]

4 Adaptation must leave room for aspiration

An enduring commitment, a project and a demanding afternoon operate on different timescales. Someone who usually wants detailed explanations may need brevity while boarding a train. That temporary request should not silently rewrite their general preference. Past behaviour also leaves out what someone wants to become. In a preregistered experiment involving 6,488 participants, Khambatta and colleagues tested article recommendations tailored to actual or ideal preferences. Targeting ideals produced somewhat fewer clicks but improved participants’ reported benefit and sense that their time was well spent. This is evidence from a recommendation experiment, not a longitudinal test of a personal assistant. [3] The implication I draw is that engagement can miss value the person cares about. A hypothetical learner preparing to present publicly might want progressively harder rehearsal, despite usually choosing easier tasks. An assistant should support that stated ambition while checking whether the challenge remains useful. The person chooses the direction.

5 The delivery of help has a cost

Timing, presentation and explanation affect the burden of assistance. In the travel example, the assistant might assemble a wardrobe around the itinerary and surface only the two compromises requiring attention. A changed evening plan should not require rebuilding the request. Amershi and colleagues’ human-AI interaction guidelines address contextual timing, efficient correction and cautious adaptation. They offer design guidance rather than proof of this proposed experience. [4] The practical question is whether the person spends less effort reaching a useful decision. A polished recommendation that creates more checking and repair may fail that test.

6 History must earn its place

An assistant shapes the evidence it collects. Its suggestions determine which options the person sees. A narrow range can generate consistent choices that reflect the system’s limitations. Acceptance may indicate preference, convenience or fatigue. Silence may justify no update at all. Useful exploration could occasionally offer a relevant alternative while preserving the familiar option. Questions should earn their cost in attention. For our traveller, whether a garment was kept, worn and comfortable would tell us more than a click, but the person decides how much feedback to provide. The strongest evaluation would compare assistance with accumulated history against the same capable system without that history, across repeated decisions. Both should receive the current request, use comparable tools and face comparable tasks. Otherwise, a stronger model or easier assignment could masquerade as the value of memory. Measure outcomes the person cares about, time spent explaining and correcting, whether corrections persist, and the burden of supervision. Include changed circumstances where old context could actively mislead. Does history improve the result after accounting for the work required to maintain it? That is a proposed evaluation standard, not a result reported here. Evidence from real users would need to show where the benefit appears, for whom, and where the cost exceeds it. Familiarity alone would not qualify as success. For example, a useful trial could track repeated packing decisions and record avoidable purchases, missing essentials and preparation time, alongside the traveller’s own assessment. It should also count the time spent supplying history. A system that saves ten minutes at departure but requires an hour of repeated correction has not established the benefit claimed here.

7 Portability must preserve meaning

A useful relationship accumulates explanations, corrections and experience. Leaving can become costly if those contributions must be reconstructed. An export can preserve records while losing the distinctions that made them useful. I would want portable context to distinguish three kinds of claim: Stated: “I want fewer wardrobe decisions during busy workweeks.” Preserve the person’s wording, when they said it and the scope they gave it. Observed: the person selected simple outfits on three work trips. Preserve the events, available options and circumstances. Selection establishes behaviour, not its reason. Inferred: the person may prefer minimalist clothing generally. Preserve that as a tentative interpretation, with supporting evidence, uncertainty and competing explanations. Do not promote it into a stated preference during transfer. Each record should carry provenance: where it came from, when it was recorded, relevant conditions and later corrections. An inference should remain linked to its evidence, so a correction can prompt review of dependent conclusions. The receiving system also needs the person’s permitted uses; importing context does not automatically authorize actions. Relevance requires a procedure. Before using an imported claim, compare its original conditions with the present decision, look for newer or conflicting evidence, and ask only if the unresolved difference could change the outcome. An old work-trip pattern might inform packing for a similar trip. It should not settle a holiday wardrobe or a new job’s expectations. This is a proposed portability standard. A receiving system should be able to leave a claim unused when its evidence is missing or its scope is unclear. Tags alone cannot guarantee good interpretation, and equivalent performance elsewhere remains unproven. Exporting a copy also does not demonstrate deletion at the original provider. The aim is to preserve enough evidence and limits for another system to reassess the context responsibly, without making the person explain everything again.

8 Usefulness must remain open to challenge

What makes personal intelligence worth trusting is a record of better assistance under changing conditions. It should carry corrections forward, recognize when experience no longer applies and help the person pursue goals that history alone would not predict. At Lumina Horizon, that is the opportunity we see. The next step is to test whether continuity produces enough benefit to justify its demands on the person. Memory volume, fluent explanations and a feeling of familiarity cannot answer that question. An intelligence grows with someone when its help improves as their life changes, and the improvement can be shown. That is the standard I would ask personal AI to meet.

Sources & further reading

  1. [1] Patrick Lewis et al. “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” NeurIPS, 2020.
  2. [2] Di Wu et al. “LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory.” ICLR, 2025.
  3. [3] Poruz Khambatta, Shwetha Mariadassou, Joshua Morris and S. Christian Wheeler. “Tailoring recommendation algorithms to ideal preferences makes users better off.” Scientific Reports, 2023.
  4. [4] Saleema Amershi et al. “Guidelines for Human-AI Interaction.” CHI, 2019.
Rights & provenance
Author
Founder Reflection by Kam Oomarow, Lumina Horizon Inc.
Publisher / rightsholder
Lumina Horizon Inc.
Version
shortened-reviewed-edition
Publication
2026-11-04T04:30:38.000Z
Canonical public URL
https://luminahorizongroup.com/ideas/series/the-last-ai-that-knows-you/04-the-memory-that-belongs-to-you
Content fingerprint
38865d5d28a973fe1b231f43d5b0099a86e39173e33d6bb405ebaacc097e6374
Rights
© 2026 Lumina Horizon Inc. All rights reserved. Share the canonical link.

Stay connected with Lumina Horizon

A Reader account is optional. Manage your communication preferences, with updates as a separate choice.

Create a Reader account →