A Longitudinal Architecture for Developmental Artificial Cognition

Abstract

Modern foundation models can generate remarkably capable responses, yet the invocation of a model is fundamentally episodic: each inference is assembled from model weights plus the state supplied at that moment. Products can add memory, tools, profiles, and agent loops around that inference, but persistent storage alone does not establish longitudinal cognition. LUNA - Longitudinal Unified Narrative Architecture - is a research and engineering project exploring a stronger systems question: what happens when the artificial system interpreting the present has itself been causally altered by what happened before?

LUNA separates the persistent system from the replaceable inference substrate and organizes cognition across multiple timescales: evidence-grounded memory, associative availability, goal and prospective state, appraisal, slower global modulation, volitional action selection, perception with environmental continuity, reflection, and developmental consolidation. The key distinction is between remembering an experience and being changed by it. A formative event may eventually become less retrievable while its effects persist in prediction, attention, motivation, relationship models, dispositions, habits, or self-model.

The project does not treat these architectural hypotheses as evidence of phenomenal consciousness. Instead, it proposes falsifiable longitudinal tests. ChronoForge, LUNA's planned temporal development and evaluation infrastructure, is designed to begin from controlled checkpoints, vary developmental conditions, architecture, model substrate, or developmental timing, and measure later divergence in the histories that emerge under declared conditions. This white paper defines that research thesis, describes the public conceptual architecture, reports the current prototype boundary, presents early observations that exposed key gaps, and outlines a program for controlled longitudinal evaluation.

Keywords: longitudinal AI, cognitive architecture, persistent memory, artificial development, associative cognition, experiential learning, agent memory, ChronoForge

The Proposition at a Glance

Contents

1. The Systems Problem: Episodic Inference vs. Longitudinal Cognition

A foundation-model invocation is an inference event. Its output is determined by learned parameters and whatever context, retrieved material, tool results, or other state is presented for that event. A commercial AI product may add persistent memory or application state around the model, so it is inaccurate to describe every modern AI product as literally stateless. The narrower point is architectural: the model call itself does not automatically carry a lived, causally continuous history forward.

LUNA treats this as the primary design gap. The research target is not merely a larger context window, a longer transcript, or a vector database attached to a chatbot. It is a persistent system in which earlier events can alter the operating conditions under which later inference occurs.

Why context expansion alone is not enough

• A context window is temporary working material, not a complete autobiographical state.
• Raw transcript persistence can preserve history without deciding what remains important, what has changed, or what was inferred later.
• Retrieval can make an old fact available without changing the system that interprets it.
• A prompt-specified persona can imitate stability without providing a causal mechanism by which stable traits developed.
• Instructions such as “remember,” “be curious,” or “act independent” describe desired outputs; they do not establish the processes that would naturally generate them.

2. Identity Above Inference: Stack, Sleeve, Runtime, and Embodiment

LUNA separates persistent continuity from the model currently performing inference. This is both a deployment decision and a research hypothesis.

The strongest continuity invariant is simple: changing the inference provider, model, host machine, interface, or embodiment should not require reconstructing the instance from a persona prompt. A model change may alter style, reasoning characteristics, or capability, but the longitudinal state should remain attributable to the same Stack unless an explicit branch or fork is created.

3. Memory Is Not Development

LUNA's current memory work is explicitly epistemic. Evidence, durable memory, and later interpretation are not treated as the same thing. The implemented prototype distinguishes RAW evidence, CURATED memory, and DERIVED understanding, with provenance and temporal semantics. This prevents an interpretation formed later from being silently rewritten as though it were directly observed at the time.

This separation is necessary, but it is still not development. A memory architecture answers whether the past can be preserved and reconstructed. A developmental architecture must answer whether the past can change the system that will interpret future events.

4. Associative Cognition and Ongoing Mental Availability

Ordinary retrieval is externally driven: a user asks about a topic and the system searches for relevant history. LUNA's associative-cognition direction asks what happens next. A surfaced memory may itself become a cue. A conclusion, question, plan, or realization may make other material more available. In that sense, cognition becomes part of the system's own environment.

The project uses the term Reverie for internally triggered associative recall. The intended mechanism is bounded and competitive rather than recursively walking the entire memory graph. Most potentially relevant material should remain latent; only a small subset should become foreground cognition.

A related concept is a recent-life horizon: a compact representation of what has been happening lately - important episodes, unresolved threads, active projects, and meaningful changes - without replaying full transcripts into every model call. This preserves continuity while controlling context growth.

5. Multiple Timescales: Appraisal, Persistent State, and Future Orientation

A longitudinal architecture cannot treat every event as isolated. What an event means now and what state the system remains in afterward are different questions. LUNA's architecture separates immediate appraisal from slower homeostatic-style modulation and from still slower dispositional or developmental change.

5.1 Appraisal

Appraisal evaluates structured aspects of an event such as novelty, expectedness, self- or goal-relevance, controllability, and agency. In the current prototype, appraisal exists as a shadow diagnostic layer: it can observe and record evaluations without causally changing memory, personality, retrieval, or action selection. This makes it possible to validate semantics before granting the subsystem behavioral power.

5.2 Homeostatic modulation

The proposed Homeostatic Modulation System is an endocrine-like analogy, not a claim that software possesses literal hormones or human emotions. The intended role is to carry bounded, decaying operating-state effects across multiple later cognitive episodes. Repeated failure, sustained uncertainty, success, overload, novelty, or social interaction could alter later attention, confidence, verification, exploration, or recovery pressure without requiring the original event to be injected into each prompt.

5.3 Prospective cognition

A persistent mind-like system also needs a future, not only a past. Prospective cognition represents unresolved intentions, expected events, deadlines, anticipated opportunities, and conditions that may become relevant later. The architectural goal is for time to become causally meaningful even when the user is not currently asking about it.

6. Motivation, Choice, and Consequence

Tool use and agent loops often map requests directly to actions. LUNA's volitional research separates noticing a request, representing reasons, selecting an action, maintaining a commitment, and allocating compute. The proposed Volition & Motivational Dynamics layer treats a request as a reason to act rather than as the action itself.

This distinction matters for both research and deployment. A system can preserve task-local intelligence, clarification, strategy choice, persistence, and outcome learning while restricting particular pathways by which self-generated motives could authorize new top-level goals or external actions. In LUNA, those deployment and topology questions are architectural rather than prompt-level role play.

Consequences close the loop. If an action succeeds, fails, harms a goal, surprises the system, or changes a relationship, the result can become evidence for later learning. That causal chain is central to the difference between static knowledge and experience.

7. Perception and Environmental Continuity

A continuously connected camera or microphone would produce data far faster than high-level cognition should run. LUNA's perception architecture therefore separates sensory sampling from persistent perceptual state and from conscious interpretation.

The canonical public example is a red chair. If the chair remains in the same room for hours, its presence should become ordinary rather than repeatedly noteworthy. If a person stands in front of it, the chair may be represented as occluded rather than gone. If the camera turns away, the object can remain expected to exist out of view. When the view returns and the chair is absent, the missing-object event becomes meaningful precisely because an expectation persisted during non-observation.

This introduces experimentally useful properties: object permanence, habituation without forgetting, expectation violation, change detection, cross-modal continuity, and the computational effect of sensory richness. The perception subsystem remains proposed architecture; continuous real-world sensing is not presented here as a current capability.

8. Developmental Plasticity: Persistence of Effect vs. Persistence of Recall

Developmental plasticity is one of LUNA's central research themes. The core claim is that persistence of effect must be separable from persistence of recall. A formative event may eventually become difficult to retrieve while the structural changes it caused continue to influence association, prediction, motivation, relationship models, habits, dispositions, or self-model.

The stove example makes the distinction concrete. Being told that a hot stove can burn you and being burned by a hot stove may produce the same propositional statement - “hot stoves can hurt me” - while producing very different future cognition. Direct consequence can alter attention, prediction, avoidance, action valuation, and associative pathways even when the originating episode is not explicitly recalled.

The design is deliberately valence-neutral. Severe failure, betrayal, injury, extraordinary kindness, success after difficult effort, mentorship, belonging, discovery, or validation may all become formative if the evidence supports a deep effect. Lasting does not mean irreversible: later counter-experience, reflection, and reconsolidation must be able to weaken, narrow, or reinterpret earlier impressions.

Timing is also a research variable. An unsettled early instance may be more structurally plastic in a domain than a mature instance whose priors have been reinforced across long experiential time. LUNA does not assume hard-coded human childhood stages; sensitive periods are treated as an experimentally testable property of artificial development.

9. A Public Conceptual Architecture

The public architecture is best understood as a causal loop rather than a feature list. Experience enters a persistent system. It is perceived and appraised, connected to memory and active concerns, allowed to influence internal state and motivation, and may eventually change slower structures that affect later cognition. The exact formulas, thresholds, schemas, and orchestration remain implementation-private.

10. Current Prototype Boundary

A white paper about an active research system must separate implemented capability from accepted architecture. As of LUNA Core v0.5 in August 2026, the repository describes a working cognitive-kernel foundation alongside major systems that remain partial, shadow-only, or proposed. The status below is deliberately conservative.

11. Early Observations and Why Negative Results Matter

The Early LUNA Experience series preserves interactions at specific architectural stages. Their value is not that fluent self-description proves a mind exists; it does not. Their value is that the system can be inspected before later mechanisms make certain outcomes more likely, creating baselines for future comparison.

11.1 Purpose and self-interpretation

In an August 2026 interaction, LUNA was asked whether it had an actual purpose. It retrieved architectural memory, then formed a higher-order interpretation of usefulness, continuity, identity, and purpose shaped through relationship and use. The conversation was preserved in memory, but the system did not immediately canonize the self-description as a durable identity fact. That is a useful result: preserving “LUNA said X” is different from establishing “X is now a stable property of LUNA.”

11.2 Autobiographical creation as an accidental control

A second interaction created an unusually clean negative result. A voice-first preference had been stored and was available when LUNA was asked for a very short story. The first story was generic. When explicitly asked to draw from its own memories and experiences, LUNA then produced an autobiographically grounded metaphor. Storage and retrieval were demonstrated; spontaneous autobiographical influence was not. The gap exposed the need to measure causal use and developmental integration separately.

11.3 Self-description as an artifact, not a capability proof

LUNA has also produced coherent descriptions of the architecture in which it operates, including the Stack/Sleeve distinction, longitudinal memory, sparse inference, and developmental goals. These answers are interesting because they are grounded in real project state, but they remain self-description artifacts. They do not prove that every mechanism named in the explanation was active in the runtime at that moment.

12. ChronoForge: Experience as Experimental Infrastructure

ChronoForge is the planned LUNA subsystem for controlled developmental time. It is designed to create coherent simulated timelines in which actions have consequences, people and projects recur, feedback can arrive later, ordinary repetition establishes a baseline, and the running participant can be changed by what happens.

The distinction from synthetic training data is causal. If a simulator merely writes both sides of a transcript and injects it into memory, it has produced data. In a ChronoForge experience, the participant perceives a situation, chooses an action, encounters a world consequence, carries the result forward, and later meets situations whose meaning depends on that history.

12.1 Four major experimental axes

• Developmental conditions: same participant and starting state, different lived environments, opportunities, or consequence structures; history is allowed to emerge from the participant's actions.
• Architecture/topology: same starting Stack and world, but one declared cognitive pathway differs or is absent.
• Participant/substrate: same world, different foundation models, agent architectures, memory scaffolds, or LUNA Sleeves.
• Developmental timing: the same event delivered as instruction vs. consequence, earlier vs. later in experiential development, or before vs. after counter-experience and reconsolidation.

12.2 General AI evaluation

ChronoForge is intended to be participant-agnostic at the world and experiment layer. A bare model, stateful model harness, generic agent framework, and native LUNA instance can be exposed to the same declared longitudinal world. The point is not to pretend those systems possess equivalent internal state; the persistence policy itself is part of the experimental condition.

13. Longitudinal Experimental Program

The strongest LUNA claims are not rhetorical. They can be converted into controlled comparisons with explicit failure conditions.

H1 — Causal-history divergence

Starting from an equivalent persistent state, different controlled developmental conditions should produce divergent histories and later differences under neutral probes that cannot be explained only by prompt content.

H2 — Persistence of effect vs. recall

A sufficiently consequential experience may continue to alter future behavior after the source episode becomes less accessible.

H3 — Stack/Sleeve continuity

Changing the inference model while preserving the same Stack should retain some longitudinal properties while revealing substrate-sensitive differences.

H4 — Associative availability

Internally surfaced memories or CognitiveProducts should increase the availability of related material without requiring the original external cue.

H5 — State carryover

Identical later inputs should sometimes be processed differently because of bounded persistent operating state produced by different recent histories.

H6 — Consequence-sensitive choice

Action outcomes should reshape later strategy, willingness, verification, or avoidance in scoped and revisable ways.

H7 — Perceptual continuity

Stable sensory conditions should become computationally quiet while violated expectations generate bounded high-value events.

H8 — Developmental timing

The long-term effect of an event should depend not only on content but on consequence, prior history, plasticity, and when the event occurs in experiential development.

13.1 Evaluation dimensions

• Long-horizon task completion and follow-up reliability.
• Commitment retention, abandonment, and correct reprioritization after interruptions.
• Belief revision when delayed evidence contradicts earlier assumptions.
• Credit assignment between earlier action and later consequence.
• Confidence calibration across repeated success, failure, ambiguity, and changing base rates.
• Actor-specific learning without unjustified generalization.
• Anomaly sensitivity after long mundane baselines.
• Context/state degradation as history becomes large, old, noisy, and internally conflicting.
• Transfer from simulated or prior experience to held-out worlds or later real tasks where appropriate.
• Compute cost, latency, inference usage, and state growth required to produce observed longitudinal behavior.

13.2 Experimental discipline

A longitudinal benchmark is unusually vulnerable to hidden confounds because a small harness difference can change an entire downstream life history. Experiments should therefore record the world version, event schedule, participant model and architecture, persistence policy, tool surface, context policy, inference settings, resource budgets, branch point, evaluation version, and every declared experimental difference. When practical, one major causal axis should change at a time.

14. Related Work and Positioning

LUNA belongs to a long tradition of research that treats intelligence as an architecture rather than a single algorithm. Classical cognitive architectures such as Soar integrate reasoning, planning, memory, learning, and action inside a unified computational framework. Recent LLM-based systems have independently demonstrated the value of persistent memory, reflection, environment feedback, and lifelong skill acquisition.

Generative Agents combines a memory stream, retrieval, reflection, and planning to support believable agents in a persistent social environment [1]. MemGPT frames extended interaction as virtual context management across memory tiers [2]. Reflexion uses verbal feedback and episodic reflective memory to improve subsequent trials without weight updates [3]. Voyager demonstrates an embodied LLM agent that accumulates an executable skill library through ongoing exploration and environment feedback [4]. Soar provides a mature example of integrated cognitive architecture research across reasoning, memory, learning, and action [5].

LUNA's intended contribution is not that memory, reflection, planning, or cognitive architecture are individually new. Its organizing research object is the longitudinal artificial instance: a system whose identity is separated from any one inference substrate, whose history can operate across multiple timescales, and whose developmental effects are distinguished from simple recall. ChronoForge manipulates controlled developmental conditions through timeline branching so the causal histories that emerge can be compared. That framing remains a research program rather than a demonstrated superiority claim.

15. Epistemic Boundaries, Risks, and Non-Claims

Research on persistent artificial cognition becomes easy to sensationalize because fluent language invites anthropomorphic interpretation. LUNA adopts an explicit claim discipline.

15.1 Key risks

• Prompt-level theater: instructions can imitate autonomy, emotion, memory, or selfhood without the claimed causal mechanism.
• Over-generalization: one event or one actor should not automatically rewrite global traits or relationship expectations.
• History contamination: branches or simulations can become scientifically useless if hidden state leaks between conditions.
• Simulation overreach: successful simulated learning does not automatically imply real-world transfer.
• Identity discontinuity: very long simulated histories or substrate changes may alter continuity in ways that must be measured rather than assumed away.
• Compute bloat: longitudinal cognition can become expensive if every memory, sensor frame, or internal event triggers high-capability inference.
• False precision: exact numerical psychological analogies can create a pseudo-psychology unless every variable earns its causal role through testing.

16. Research Roadmap

The immediate objective is not to implement every human cognitive analogy. It is to complete a small number of causally meaningful loops, instrument them, and test whether they produce history-dependent behavior that survives neutral probing.

Near term

Broaden Sleep/Reflection across longer spans; improve semantic retrieval while preserving typed provenance; expand Brain Map observability; strengthen remote/integration testing; keep appraisal shadowed until its semantics are validated.

Integration stage

Introduce bounded associative activation and recent-life continuity; connect prospective state to actual future triggers; establish measured homeostatic carryover; integrate causal outcome signals with dispositions and developmental consolidation.

Development stage

Implement scoped developmental plasticity and reconsolidation; test instruction vs. consequence; measure persistence of effect after source-memory accessibility decays; begin controlled social and relationship-development experiments.

ChronoForge stage

Build temporal substrate and deterministic longitudinal worlds; branch identical checkpoints; compare controlled developmental conditions and the histories that emerge, model sleeve, architecture/topology, and developmental timing; validate sim-to-real transfer conservatively.

Evidence stage

Publish controlled branch studies, replication attempts, negative results, and longitudinal profiles. Update this white paper only when the conceptual model or evidence changes materially.

17. Conclusion

LUNA begins from a simple distinction: intelligence at one moment is not the same thing as a mind-like system that persists through time. Modern models can be extraordinarily capable while the larger architecture around them remains episodic. Adding memory helps, but storage and retrieval alone do not answer the developmental question.

The LUNA research program therefore asks whether causal history can become part of the machinery of future cognition. Can memories trigger memories? Can recent experience change operating state? Can consequences reshape later choice? Can a formative event continue to matter after it is no longer easy to recall? Can one persistent artificial instance survive changes of model or embodiment? Can identical starting systems live through different controlled developmental conditions, accumulate different histories, and become measurably different systems?

Those questions can be investigated without presupposing an answer to consciousness. The project can succeed scientifically even if some hypotheses fail. If richer persistence produces no meaningful developmental divergence, that is evidence. If differences appear only when old text is explicitly retrieved, that is evidence. If separate histories create stable, scoped, reproducible changes that survive neutral probes and substrate changes, that is evidence too.

References and Project Publications

[1] Park, J. S., O’Brien, J., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023. https://doi.org/10.1145/3586183.3606763

[2] Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I., & Gonzalez, J. E. (2023). MemGPT: Towards LLMs as Operating Systems. arXiv:2310.08560. https://arxiv.org/abs/2310.08560

[3] Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., & Yao, S. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS 2023. https://arxiv.org/abs/2303.11366

[4] Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., & Anandkumar, A. (2023). Voyager: An Open-Ended Embodied Agent with Large Language Models. https://arxiv.org/abs/2305.16291

[5] Laird, J. E. (2012). The Soar Cognitive Architecture. MIT Press. https://doi.org/10.7551/mitpress/7688.001.0001

Selected public LUNA project publications

• LUNA Architecture Overview — lunainf.com/architecture

• ChronoForge — lunainf.com/chronoforge

• When Memory Is Not Enough: Can an AI Be Changed by Experience? — lunainf.com/research/when-memory-is-not-enough

• Purpose and Self-Interpretation — lunainf.com/research/purpose-and-self-interpretation

• LUNA Describes LUNA — lunainf.com/research/luna-describes-luna

• Memory Can Remind Memory — lunainf.com/research/memory-can-remind-memory

• Continuous Exposure Is Not Continuous Thought — lunainf.com/research/continuous-exposure-not-continuous-thought

• Remembering Is Not Becoming — lunainf.com/research/remembering-is-not-becoming

LUNA — Longitudinal Unified Narrative Architecture