Skip to content

Deep Tech & Excellence

AI that knows the person. Care that grows with them.

Nightingale PALI — Personal Adaptive Lifelong Intelligence. Private understanding. Shared progress. Across a lifetime.

SynaptixLabs research direction · HappySeniors / Nightingale

Proposed research programme · PALI is a working research name

Download the article (English PDF) (opens in a new tab)

The ambition

Nightingale proposes personal AI that develops a lasting, permissioned understanding of an individual: what matters to them, how they live, and how their support needs change. Across rehabilitation, home and community care, that understanding would help the person and their professional team anticipate needs, coordinate providers and follow through on a shared plan. We call this proposed research programme Personal Adaptive Lifelong Intelligence—PALI. Its ambition has two connected parts: learning deeply for one person, and turning appropriately authorised, privacy-tested learning into better support for others. The individual’s evidence and learned state would remain distinct from any contribution considered for shared learning. The central research task is a controller that decides whether new evidence warrants a persistent update to the person-specific adaptive neural region in PALI, and whether that update needs additional capacity. The controller would test candidates before activating a new personal version, and can instead retrieve records, consult a companion, share an eligible finding or abstain. Large models, small personal models and hybrid architectures will compete against strong conventional and retrieval-based baselines. Evidence-linked explanations, professional correction, permission controls and staged validation are central to the design. The proof must show useful continuity and transferable benefit at acceptable disclosure risk, false-alert burden and cost. Established methods provide the foundations; the mechanism’s novelty, enduring value and care outcomes remain to be demonstrated.

Figure 1. Evidence-gated personal learning architecture

Select a module.

Proposed internal architecture. Blue shared capability and amber personal capacity are interwoven in one model. Isolation and useful capacity growth require testing.

Shared promotion has a separate permission and validation process; it does not follow automatically from personal learning.

01A lifelong ambition, grounded in an ordinary care transition

The care context spans older adults and adults with disabilities. Consider a fictional older person returning home after rehabilitation. Walking has improved, but fatigue changes when they shop, a daughter’s availability has changed, and a new provider knows little about the routines that make help acceptable. A later decline might look small in each record yet matter across the whole story. A young adult with a disability starting independent living faces a parallel need: continuity while goals, assistance and providers change. Useful AI would help the person and team connect these observations, revisit the plan, find the right provider and check whether agreed support actually happened.

Here, “knows the person” means permissioned, revisable understanding grounded in sources—including what is unknown. It does not mean an all-knowing agent. The proposed system would distinguish an enduring preference from a temporary constraint, deterioration from recovery, and a professional correction from an unverified report. The person remains an active participant whose goals and permissions can change; responsible professionals retain authority over care decisions.

Personal neural capacity means adjustable model parameters or learned state allocated to an individual, updated during use and deliberately retained between interactions. It would complement a dated evidence record and a stable historical reference, not replace either. “Across a lifetime” is a design horizon for continuity across changing services, rather than a trial duration or a commitment to retain every private fact forever. The challenge is to make that continuity useful in a practical care workflow, not simply to remember more.

02What existing systems already learn—and what remains unproven

Personalisation has substantial foundations

OPPU combines personal trainable modules with retrieved history [1]. PROPER studies adaptation across population, group and individual levels [2]. Parameter-efficient fine-tuning (PEFT), including low-rank adaptation (LoRA), changes a relatively small set of parameters rather than retraining an entire model. These are direct precedents for personal capacity within a shared system. LINEUP goes further in questioning how much private capacity is needed: its shared factor bank and compact user code challenge the assumption that everyone needs a substantial independent adapter [4].

Profile-to-PEFT generates adapter weights from a profile using a hypernetwork [3]; related on-device work generates LoRA parameters from context [12]. Generated weights come from forward computation, which differs from gradient training during an encounter. Such approaches might offer an efficient personal companion, but they should compete with learned persistent states rather than be relabelled as the same mechanism. Retrieval is also a strong baseline, especially when histories are sparse [9]. Personalisation can worsen previously correct answers, so preference fit and factual reliability need separate evaluation [10].

A history, an adaptive state and a shared strategy are different things

Test-Time Training layers update a trainable hidden-state model while processing a sequence [5]. Dynamically Expandable Networks selectively retrain and add capacity as tasks arrive [6], while Memory Layers at Scale explores sparse trainable memory [7]. None automatically supplies an accountable personal state that persists across sessions and years. A retention policy, update rule, version history and correction mechanism must make persistence deliberate. Expanding a network is an established technique, not by itself a scientific contribution.

Agent-memory research also addresses the relationship between shared and individual learning. HiPS uses global shared strategies and user-specific adaptive strategies, making it direct prior art for a two-level learning ambition [13]. THEANINE links memories through timelines to interpret change [15]. Memora evaluates the cost of obsolete memories as well as useful recall over weeks and months [14]. PALI must compare with these mechanisms instead of treating historical continuity or global/personal separation as undiscovered ideas.

Prompts, retrieval systems and shared knowledge resources can benefit many people. Retrieving one person’s record does not, by itself, update shared model weights; broader improvement may instead come from a better shared prompt, a vetted knowledge item, an updated strategy or training. A Nature Health perspective connects continuity, coherence, adaptation and agency in longitudinal health AI [8], and MedCL-Bench examines sequential biomedical learning [11]. These foundations motivate the setting, but do not establish PALI’s mechanism or care benefit. The research task is a specific, falsifiable combination under care constraints.

03Two learning loops, with different permission to change

Figure 1 shows the internal PALI architecture. A fixed shared LLM and a person-specific reconfigurable neural network contribute to the same inference. Dated evidence stays in a separate record. A personal network update changes learned parameters or connections, which influence future inference for that person. Permitted new context may come from the person, an observed cue or a professional correction; its source, date and uncertainty remain explicit. A controller would choose among retrieving existing evidence, proposing a personal-state update, allocating capacity, consulting a companion model, or abstaining. These are decisions to define and learn, not interchangeable labels for an API call. A new observation may be worth retaining without justifying any neural update; a changed preference may justify an interpretation update without changing a historical fact.

The controller would compare an update candidate with a stable reference and the current reviewed version. A verified correction, delayed outcome or conflicting report could support promotion, require revision, or leave the previous state in use. Figure 2 illustrates the proposed acceptance rule. First check evidence provenance, permission and comparable context. Then use permitted earlier visits to train candidate personal parameters in isolation. Hold later visits out of training and compare the current model with retrieval, an update within existing capacity and an expanded candidate. Activate a larger personal network only if it improves recognition enough to justify its cost and meets predefined limits for false alerts, retention and disclosure. Keep the original evidence and mobility baseline, and retain the previous model version for rollback. The blue shared-network geometry remains fixed; the added units are schematic, not measured parameter counts. The two figures illustrate the personal loop; the shared-learning loop described here is a separate permission and validation process. Learning during use does not imply that every change becomes immediately available to support care.

The shared loop would ask a different question: is there an eligible contribution that could improve support for others? Possible contributions include an evaluated learning rule, a generalised workflow insight, a privacy-protected aggregate or a candidate model update. Eligibility would depend on the permission attached to the source and the intended use, disclosure tests, and evidence that the contribution generalises. A person’s record or learned state would not be silently copied into a common pool merely because the system found it useful.

Shared promotion would have its own version, evaluation set and rollback decision. Candidate improvements would first be tested on held-out people, groups and settings, including those unlike the originating contributor. Benefit to one person must not conceal harm to another. The personal controller could later use an approved shared version, but that dependency would remain traceable. An organisation could also run with no shared promotion; that is an important scientific comparator and a legitimate permission configuration, not a broken version of the design.

The proposed scientific contribution is the operational mechanism connecting these decisions: what evidence permits an update, what justifies more capacity, which contribution may cross a boundary, and when each change should be withheld. We would test whether the mechanism improves useful recognition and coordination while retaining correction, historical comparison and control. The general idea of personal plus shared learning is already researched; the value must come from a defined mechanism and reproducible behavior.

04Large models, private companions and boundaries that must be tested

An enterprise-scale shared model could provide broad capabilities while holding separately addressed personal state. A smaller private companion could keep the changing personal representation and consult a general model when needed. A hybrid could split tasks, with local continuity and selective access to a larger model’s capabilities. “Private” describes the intended access and learning boundary, not a privacy guarantee from model size, device location or a coloured region in a diagram. Internal parameter allocation requires control of the model architecture and training access; with only a hosted inference API, persistent personal learning would need supported adaptation or an external companion. The trade-offs include latency, storage, specialist competence, updating costs and governance.

RouteLLM establishes quality–cost routing between models as relevant prior art [18]. PALI adds questions about what persistent state travels with a request, which model may learn from it, and whether routing changes reliability. We would compare a large-model-only system, a small-model-only system and matched hybrids, with equal access to permitted information. Permissioned organisational silos, centralised processing and federated learning would be alternative configurations to test rather than a predetermined winner.

Local records alone do not make federated updates safe. A July 2026 preprint studies an active malicious server that alters the distributed architecture and reconstructs radiology text through gradient inversion in its experimental setup [17]. MEXTRA demonstrates private-memory extraction through prompts in evaluated agents [16]. These findings motivate explicit threat models around prompts, retrieval, weights, gradients, logs and update transport. They do not imply that every architecture leaks in the same way or that moving data off-device is the only risk.

For distributed updates, we would test secure aggregation and person-level differential privacy, reporting the privacy budget, extraction risk, transfer benefit and compute cost; neither technique alone makes shared learning automatically private.

The implementation would need purpose-specific consent and permissions, identity and key management, access enforcement, audit trails and controlled retention. Revoking access can block future use but does not automatically erase influence already encoded in model weights; unlearning or retraining must be tested with declared limits. For applicable US deployments, HIPAA-related safeguards would be design objectives assessed against the actual entities, data flows, business-associate relationships and agreements. HHS defines that applicability; this research proposal is not a claim of HIPAA compliance [24].

05Accelerating the research without bypassing evidence

A second research ambition is to improve how candidate mechanisms are developed. AI could propose update algorithms, routing policies, small-model architectures, simulation scenarios or evaluator code. An evolutionary search would generate variants, measure them and retain useful candidates. AlphaEvolve offers a concrete precedent for generated code evaluated iteratively on computational and mathematical tasks [20]. It supports a bounded research workflow, not a promise of exponential progress or a clinical system that rewrites itself in production.

People would set the objectives and acceptance limits before a search. Candidate code would run in an isolated research environment, with fixed baselines, versioned experiments and held-out tests kept separate from the search feedback. Proposed evaluators would themselves need review and validation; the same model must not silently redefine success to reward its own output. Promotion would require reproducibility, human review and rollback. Search efficiency is an outcome to measure alongside reliability, not an assumed acceleration factor.

Simulation offers two distinct tools. NVIDIA’s Omniverse-based Isaac Sim and Cosmos world models could support candidate studies of physical environments, sensing and robotics—for example, a proposed home or rehabilitation scenario with mobility aids, changed lighting or sensor failures. The cited Isaac for Healthcare work concerns surgical and interventional device–anatomy simulation [23]; it does not validate long-term care scenarios. Cosmos 3 studies physical-world models and synthetic data, and some synthetic-data ablations reduce human-domain performance, motivating simulation-to-real validation [22]. Separately, synthetic longitudinal care scenarios could encode changing goals, missing reports, delayed interventions and service transitions to stress-test a controller. Neither tool is a validated digital twin of an individual’s health or lived experience.

V-JEPA 2.1 is a related frontier probe: a Joint Embedding Predictive Architecture learns representations by predicting latent features, with research on dense video representations and potential use in physical perception and planning [21]. It may be relevant where a partner brings physical sensing or robotics. That is a different role from clinical reasoning or provider coordination. Representation learning, synthetic scenarios and physical simulation should enter the programme only where a defined experiment tests their contribution, followed by separate checks against real observations.

06Explanations that can be checked

A professional should be able to inspect the observations behind a suggestion: source, date, permissions, uncertainty, competing interpretations and what changed since the previous version. The system should distinguish a retrieved fact from an inferred pattern, show a correction’s effect, and explain why it asked for review or declined to act. This is an evidence and change trace that can support responsible work. It also gives the person a practical way to challenge an interpretation instead of accepting a fluent narrative as authority.

That trace is not full neural transparency. A 2025 medical-question study finds that model explanations can misattribute how evidence influenced an answer [19]. An apparently plausible rationale therefore needs its own tests: remove or alter a cited observation, introduce a conflicting source, compare alternative explanations, and check whether the result changes as claimed. Explanation faithfulness, predictive accuracy and professional usefulness are separate outcomes. Where an internal causal account cannot be verified, the interface should expose the observable evidence and limits rather than invent a precise story about hidden reasoning.

07Proving the advantage—and exposing the limits

Start with conventional care workflows and explicit rules, not only newer AI models. A well-designed checklist, provider handover process or rules-based alert may solve the target problem more reliably. Compare those with retrieval over dated records, timeline memory, fixed adapters, compact personal codes and global/personal strategies. Match available history and resource budgets. Vary model size, internal versus companion capacity, centralised versus federated or no-sharing configurations, and shared-promotion permission. Each comparison should isolate a decision the proposed controller actually makes.

Development and evaluation must separate people, organisations where feasible, and time. At each decision, expose only information available then; keep later observations and corrections for testing. Include cold starts, contradictory reports, outdated memories, gradual deterioration, recovery, changed goals and transitions between services. Transfer tests should ask whether an approved shared contribution helps previously unseen people and settings, and who loses when averages improve. A retrospective gain on one dataset is not evidence of lifelong benefit.

Benchmarks would evaluate useful recognition at a fixed false-alert burden, calibration, important omissions, factual reliability and abstention. Track correction retention, forgetting of earlier useful knowledge, cross-person interference and disclosure under the declared attacks. Report latency, storage and compute with the actual model and deployment configuration. In workflow studies, examine professional review time, inappropriate escalation, completed handovers and provider follow-through. The person’s agency, accessibility and the team’s ability to correct or reject suggestions are part of practical value.

Ablations should remove the stable reference, update gate, capacity-growth rule, companion route and shared-promotion gate one at a time. Freeze the other conditions and ask which component explains a benefit. For the accelerated research process, measure whether AI-generated candidates outperform human-designed and simpler searches under the same evaluation budget. For explanations, test faithful response to evidence interventions rather than scoring fluency alone. Reproducible negative results should be useful outputs of the programme.

The stop rule is explicit: do not advance learned capacity or shared promotion if it lacks meaningful benefit over matched conventional or retrieval baselines, worsens reliability, or breaches agreed disclosure, workload or resource limits. Simulation first isolates mechanisms; governed retrospective data tests real variation; prospective assisted use requires a separate evidence and oversight gate. Progression depends on predefined criteria and responsible partners, not on an attractive simulation or a self-reported model score. Numerical efficacy targets belong in a jointly designed protocol, not an unsupported marketing promise.

Table 1. Comparators that test the value of the mechanism
ComparatorWhat it testsCritical question
Conventional rules and care workflowValue beyond process improvementDoes the AI mechanism improve on a clear checklist, alert rule or accountable handover?
Fixed model + retrieval / timeline memoryEvidence access without persistent neural updatesDoes learning beat matched access to dated, corrected and obsolete records?
Adapters, compact codes and generated weightsEfficient fixed or generated personalisationIs extra persistent capacity necessary for the task and history available?
Global/personal memory strategiesTransfer and personal continuityDoes the proposed controller improve on established two-tier learning?
Small, large and routed model systemsCapability, latency and resource allocationWhich size and routing policy earns its cost under the same reliability limits?
Centralised, federated and no-sharing variantsPermission boundaries and shared promotionDoes permitted transfer help unseen people without unacceptable disclosure or interference?
Fixed capacity versus growth/companion variantsThe contribution of adaptive allocationWhich component helps when information and evaluation budgets are matched?

08From a testable mechanism to reproducible evidence

Developing this mechanism requires complementary work in continual learning, efficient model architectures, privacy engineering, care methodology, disability-led co-design, provider workflows and independent evaluation. Scientific collaboration should begin by specifying the controller, reproducible comparators and the evidence needed for responsible use.

A shared research protocol should define data permissions, chronological evaluation, responsible review and conditions for stopping or revising the mechanism. Publishing the comparison methods, resource costs, limitations and reproducible negative results would help others assess what the research actually establishes. The aim is useful continuity for the person, supported by evidence that can be challenged and improved.

References

Selected research publications and primary technical sources from a targeted scoping review through 16 November 2026. Conference, preprint and engineering-source status are identified below. This is not a systematic review or a claim that results were independently reproduced.

  1. Zhaoxuan Tan et al. (2024). Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuningEMNLP 2024.
  2. Linhai Zhang et al. (2025). PROPER: A Progressive Learning Framework for Personalized Large Language Models with Group-Level AdaptationACL 2025.
  3. Zhaoxuan Tan et al. (2026). Instant Personalized Large Language Model Adaptation via HypernetworkACL 2026.
  4. Songyuan Sui et al. (2026). Does Every User Need a Private LoRA? Decoupling Personalization from Per-User AdaptationPreprint, October 2026.
  5. Yu Sun et al. (2025). Learning to (Learn at Test Time): RNNs with Expressive Hidden StatesICML 2025.
  6. Jaehong Yoon et al. (2018). Lifelong Learning with Dynamically Expandable NetworksICLR 2018.
  7. Vincent-Pierre Berges et al. (2025). Memory Layers at Scale2024 preprint; ICML 2025.
  8. Georgianna Lin, Rencong Jiang, Noémie Elhadad and Xuhai “Orson” Xu (2026). A framework for longitudinal health AI agentsNature Health Perspective, 2026.
  9. Alireza Salemi and Hamed Zamani (2025). Comparing Retrieval-Augmentation and Parameter-Efficient Fine-Tuning for Privacy-Preserving Personalization of Large Language Models2024 preprint; ICTIR 2025.
  10. Chimaobi Okite et al. (2025). Benchmarking and Improving LLM Robustness for Personalized GenerationEMNLP Findings 2025.
  11. Min Zeng et al. (2026). MedCL-Bench: Benchmarking stability-efficiency trade-offs and scaling in biomedical continual learningPreprint, March 2026.
  12. Sean Augenstein et al. (2026). LoRA-generating hypernetworks for efficient on-device LLM generative personalizationPreprint, September 2026.
  13. Yupeng Han et al. (2026). Learning What to Share and What to Personalize: Hierarchical Strategy Co-Evolution for Agent MemoryEMNLP 2026.
  14. Md Nayem Uddin et al. (2026). From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized AgentsFindings of ACL 2026.
  15. Kai Tzu-iunn Ong et al. (2025). Towards Lifelong Dialogue Agents via Timeline-based Memory ManagementNAACL 2025.
  16. Bo Wang et al. (2025). Unveiling Privacy Risks in LLM Agent MemoryACL 2025.
  17. Santhosh Parampottupadam et al. (2026). Privacy Leakage in Federated Learning in Radiology Reports: A Comparative Evaluation of Tokenizer-Driven Privacy RisksPreprint, July 2026.
  18. Isaac Ong et al. (2025). RouteLLM: Learning to Route LLMs from Preference DataICLR 2025.
  19. Katie Matton et al. (2025). Walk the Talk? Measuring the Faithfulness of Large Language Model ExplanationsPreprint, 2025.
  20. Alexander Novikov et al. (2025). AlphaEvolve: A coding agent for scientific and algorithmic discoveryWhite paper / preprint, 2025.
  21. Lorenzo Mur-Labadia et al. (2026). V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised LearningPreprint, March 2026.
  22. NVIDIA (2026). Cosmos 3: Omnimodal World Models for Physical AINVIDIA technical report, 2026.
  23. Cristiana Dinea et al. (2026). Developing Healthcare Robotics with GPU-Native Medical Physics SimulationNVIDIA technical blog, July 2026.

Deployment governance reference

[24] US Department of Health and Human Services, Office for Civil Rights. Covered Entities and Business Associates. Official HIPAA guidance; accessed 16 November 2026.