Agentic AI for Cultural Heritage: Embedding Risk Memory in Semantic Digital Twins

Abstract

Cultural heritage preservation increasingly relies on data-driven technologies, yet most existing systems lack the cognitive and temporal depth required to support meaningful, transparent, and policy-informed decision-making. This paper proposes a conceptual framework for memory-enabled, semantically grounded AI agents in the cultural domain, showing how the integration of the ICCROM/CCI ABC method for risk assessment into the Panoptes ontology enables the structured encoding of risk cognition over time. This structured risk memory becomes the foundation for agentic reasoning, supporting prioritization, justification, and long-term preservation planning. It is argued that this approach constitutes a principled step toward the development of Cultural Agentic AI: autonomous systems that remember, reason, and act in alignment with cultural values. Proof-of-concept simulations illustrate how memory-enabled agents can trace evolving risk patterns, trigger policy responses, and evaluate mitigation outcomes through structured, explainable reasoning.

1. Introduction

Cultural heritage assets are increasingly exposed to complex and evolving risks, including natural, anthropogenic, and systems remain largely reactive. They often fail to correlate present conditions with historical patterns, lack mechanisms for transparent and value-sensitive reasoning, and provide limited support for long-term decision-making.

Artificial Intelligence (AI) research has been shifting toward models characterized by agency and memory. These agentic AI systems maintain internal representations, reason over time, and pursue goal-oriented behaviors, properties that are essential for the proactive and context-aware preservation and management of cultural heritage.

To address this gap, a novel integration of the ICCROM/CCI ABC method for risk assessment into a newly published heritage ontology, Panoptes, is proposed. This integration enables the semantic encoding of risk memory and supports agents that can prioritize, justify, and simulate heritage-related decisions by drawing upon structured representations of past threats. The resulting framework establishes a foundation for cognitively enriched, memory-enabled AI systems aligned with the values and priorities of cultural preservation. Simulation results showcase the potential of such an integration. This study does not claim real-world validation but offers a conceptual simulation based on plausible assumptions, to evaluate the feasibility of embedding risk memory into agentic AI workflows.

2. Background and Related Work

Agentic AI has emerged as a paradigm for intelligent systems that maintain internal state, reason across time, and pursue structured goals based on environmental perception and memory. Recent studies have converged on a more structured understanding of agentic AI, distinguishing it from tool-based systems by emphasizing autonomy, internal goal management, temporal reasoning, and adaptive behavior. The notion of “machine agency” as a psychological framework for understanding human–AI interaction dynamics has been proposed [1]. Complementing this, OpenAI outlines best practices for governing agentic systems, stressing the alignment of emergent behaviors with human oversight and ethical boundaries [2]. Finally, the design and deployment of agentic architectures has been categorized, highlighting their capacity to operate under complex, evolving goals in real-world scenarios [3]. Agentic AI properties align closely with the requirements for proactive and context-sensitive cultural heritage reasoning agents. In contrast to narrow AI, agentic models emphasize autonomy, persistent internal representations, and decision-making grounded in temporal reasoning [4]. Such architectures are gaining traction in various domains, like robotics, complex planning, and autonomous systems, yet their application in structured knowledge domains like cultural heritage remains limited.

Among those limited attempts, Sun et al. [5] explored the emergence of communication and representation by developing AI Nüshu, an interactive AI installation inspired by a historically female-only script. Their work exemplifies agentic behavior in creative cultural expression, with agents iteratively forming a shared writing system from sparse symbolic input. Suh et al. [6] designed a Langchain-based chatbot agent to provide historical and curatorial information about cultural sites in Seoul. This initiative demonstrates how conversational agents can enhance public engagement with cultural narratives using city-authenticated datasets. In the domain of autonomous documentation, Giakoumidis and Anagnostopoulos [7] proposed ARM4CH, a methodology leveraging quadrupedal robots and UAVs to autonomously perform 3D reality modeling of heritage assets; this agentic system minimizes human intervention while enhancing coverage and repeatability in complex environments. Anik et al. [8] introduced a multi-agent AI framework tailored for culturally adaptive translation in underserved language communities; their architecture includes specialized agents for translation, interpretation, content synthesis, and bias mitigation, offering an end-to-end pipeline for culturally grounded linguistic services. These studies underscore the growing potential of agentic AI to support documentation, translation, monitoring, and interpretation in heritage settings, offering proactive, autonomous, and culturally sensitive digital infrastructures.

On another aspect, ontologies have become significantly valuable tools for structuring cultural heritage knowledge and enabling machine-processable reasoning. The CIDOC Conceptual Reference Model (CRM) has long established a foundational standard for modeling cultural heritage data using event-based, semantically explicit relationships [9]. It is widely adopted across digital archives and museums to ensure semantic interoperability, and it could be the longest standing standard in the domain. Building on this foundation, digital twins have been recently proposed to simulate the real-time states and dynamic conditions of physical heritage assets. While initially developed for industrial monitoring [10], digital twins are increasingly explored in cultural heritage contexts for monitoring deterioration, simulating environmental impact, and supporting predictive preservation [11]. Recent semantic extensions of digital twins have aimed to incorporate real-time data via sensor integration. The Heritage Digital Twin Ontology (HDTO) and its reactive counterpart (RHDTO) introduced mechanisms for linking sensor-derived data to ontological entities, thereby enabling feedback loops and environmental reactivity [12]. However, these frameworks often remain reactive and lack structures for memory, risk prioritization, or value-aligned reasoning.

On the front of cultural heritage risk monitoring, a number of formal methodologies have been developed. Ashley-Smith’s foundational work [13] introduced structured criteria for quantifying deterioration mechanisms and linking them to preventive strategies. The Cultural Property Risk Analysis Model (CPRAM) proposed by Waller [14] adopts a probabilistic lens, enabling prioritization based on expected loss and uncertainty. The QuiskScan method [15] provides a rapid assessment framework for identifying vulnerabilities across collections, balancing qualitative judgment and systematic evaluation. Complementary instruments such as the Preservation Needs Assessment (PNA) tool by the AICCM [16] and Rowson’s “Dollars and Sense” model [17] integrate significance, condition, and cost into multivariate conservation planning. Among these, the ICCROM/CCI ABC framework has gained particular traction for its quantitative yet operational simplicity. The ABC model defines risk through three additive components: A (frequency of occurrence), B (loss per item), and C (fraction affected), whose sum constitutes the Magnitude of Risk ($MR=A+B+C$) [18]. While widely used for physical heritage preservation and policy formulation, the ABC model has not yet been systematically incorporated into semantic, agent-based, or memory-enabled computational architectures. Unlike CPRAM or QuiskScan, the ABC model’s additive decomposition ($MR=A+B+C$) maps naturally to a semantic triple structure. CPRAM expresses risk as an expected loss metric derived from probabilistic functions over uncertainty, vulnerability, and impact; this continuous and expert-calibrated structure is difficult to encode in atomic ontological form. QuiskScan, on the other hand, applies ordinal matrices across multiple vulnerability dimensions but lacks componential definitions that align with semantic rules or traceable inference. In contrast, ABC’s discrete and interpretable components allow each risk assessment to be directly represented through structured ontology properties, queried via SPARQL, and acted upon by rule-based agents. This makes it particularly well-suited for agentic AI architectures requiring memory, modular reasoning, and explainability.

Some attempts have been made to formalize risk ontologies, such as the HeRO ontology developed within the SIRIUS project [19]. HeRO models threats, vulnerabilities, and institutional responses but remains descriptive and lacks embedded memory or simulation. Similarly, agentic AI concepts have recently begun to surface in the context of digital twins [20], though applications to ontologically structured heritage knowledge systems are still nascent. To address these limitations, the Panoptes ontology was developed within the ARGUS Horizon Europe project and was accepted for presentation at high-impact venues such as CAA2025 and IEEE-CH2025 [21,22]. It introduces a modular, multimodal, and extensible semantic model for representing heritage assets, spatiotemporal events, sensor measurements, satellite data, policy-based decisions, and decision support, towards true preventive preservation in cultural heritage. Panoptes is aligned with a suite of standards, including CIDOC CRM, and leverages ontological standards such as SOSA/SSN for sensor observation [23], OWL-Time for temporal reasoning [24], PROV-O for provenance modeling [25], and GeoSPARQL [26], and it is implemented under FAIR data principles for interoperability and reuse [27].

This paper proposes embedding the ABC risk model directly into the semantic core of Panoptes, transforming it from an advanced data model into a memory-capable cognitive structure. Version 3 of Panoptes, introduced in this paper, extends the previous public release (v2) used in prior conference publications (Panoptes version 2.0, online at https://gitlab.com/ilsp-xanthi-medialab/argus/panoptes-ontology/-/blob/main/panoptes2.owl, (accessed on 3 June 2025)) with formal constructs for cultural risk memory and agentic reasoning (Panoptes version 3.0, online at https://gitlab.com/ilsp-xanthi-medialab/argus/panoptes-ontology/-/blob/main/panoptes3.owl, (accessed on 3 June 2025)). This extension enables persistent, episodic records of risk cognition over time, supporting prioritization, justification, and simulation by cultural AI agents. These agents, grounded in the principles of agentic AI, may be capable of remembering risk trajectories, reason with evolving policies, and align actions with preservation values. In summary, this work bridges three key strands: (1) structured cultural risk assessment using the ABC method, (2) semantically grounded, interoperable modeling via Panoptes, and (3) agentic AI frameworks capable of memory, reasoning, and action. The convergence of these domains sets the stage for culturally aware AI agents that remember, simulate, and support preservation decisions over time.

3. From Structured Memory to Agentic Reasoning

This paper presents an extension of the Panoptes ontology to incorporate the ICCROM/CCI ABC risk assessment method, enabling semantic support for agent-based monitoring in cultural heritage preservation. The integration allows intelligent agents to maintain structured memory, analyze risk evolution over time, prioritize actions based on cultural value, and generate explainable decisions. The semantic model introduces a set of new classes within Panoptes, centered around the concept of RiskAssessment. Each instance of this class represents a temporally situated evaluation of risk affecting a specific heritage asset. These assessments are further decomposed into structured components: FrequencyA, LossB, and FractionC, which together determine the MagnitudeOfRisk ($MR$) score through the additive rule $MR=A+B+C$, as defined in ABC. Risk assessments are linked to instances of RiskSource (e.g., threats like humidity or earthquakes), the targeted Asset, and a timestamp (hasTime) aligned with OWL-Time. They may also reference contextual Policy elements, capturing the preservation strategy or conservation regime under which the risk evaluation occurred. By aggregating these entities, each RiskAssessment forms an episodic memory unit: a semantic bundle that captures the who, what, when, and why of a risk-related judgment. This structure supports a range of cognitive operations, from trend detection to simulation and justification.

The rationale behind extending Panoptes with ABC-specific constructs lies in enhancing its operational expressiveness and enabling agentic AI approaches. While Panoptes originally modeled static and dynamic asset conditions, policy contexts, and threat events, it lacked formalized risk cognition capabilities. The new ABC components in Panoptes (FrequencyA, LossB, FractionC) introduce a domain-independent structure to quantify cultural risk. These semantic elements, coupled with time-stamping and policy linkage, allow the ontology to function not just as an active metadata repository but also as a reasoning substrate, supporting long-term monitoring, prioritization, and explainable cultural AI backbone. The ontology extensions introduced were guided by three design goals: (1) Semantic coherence, achieved by reusing or aligning with standard vocabularies such as OWL-Time, PROV-O, SOSA/SSN, and GeoSPARQL; (2) Agent readiness, ensuring that classes and properties support the reasoning and temporal traceability required for agentic memory; and (3) Preservation domain specificity, capturing the epistemic structure of cultural risk assessment using the ABC method. For instance, the modular use of FrequencyA, LossB, and FractionC within each RiskAssessment instance allows the encoding of both the reasoning process and the outcome. Similarly, explicit linkage to Policy enables value-sensitive decision traces. All classes were integrated into the existing Panoptes namespace, preserving compatibility while extending semantic depth.

The integration of risk memory into Panoptes renders the ontology queryable via SPARQL and interpretable by reasoning engines. Agents can retrieve temporal sequences of $MR$ values for given assets, rank threats, or simulated future scenarios. An example SPARQL query for retrieving high-$MR$ assets over the last decade may read as shown in Listing 1.

Listing 1. Risk query over time.

SELECT ?asset ?riskSource ?mr WHERE {

?ra a :RiskAssessment ;

:assessesAsset ?asset ;

:hasRiskSource ?riskSource ;

:hasMR ?mr ;

:hasDate ?d .

FILTER(?d > "2015-01-01"^^xsd:date)

}

ORDER BY DESC(?mr)

Such queries can be embedded into agent behaviors or wrapped in middleware systems (e.g., RDF4J, Apache Jena) that support real-time or scheduled evaluations. Semantic reasoning engines, including OWL-DL reasoners or rule-based tools like SWRL or SHACL, further enhance this architecture by enabling agents to detect priority threat trajectories, infer missing policy interventions, or flag implicit violations based on $MR$ thresholds.

The reasoning framework leverages two distinct but interrelated memory modalities: (a) Episodic memory, composed of time-stamped RiskAssessment entries, supports pattern recognition, trend analysis, and the construction of risk narratives; (b) Semantic memory, which includes ontological definitions of threats, asset typologies, sensor types, and policy frameworks, supports classification, generalization, and analogical reasoning. This dual-layer memory system allows agents to not only recognize what has happened but also to interpret it within structured cultural knowledge and institutional goals.

Agentic behavior is guided by declarative rules encoded as thresholds or triggers. For instance:

If $MR>6$ and rising in the last 3 years, and no policy is active, then flag the threat as urgent.

These rules may be defined using semantic rule languages or encoded as procedural workflows within agent architectures (e.g., JADE, Protege+Jess, or web-based RDF platforms). Each recommendation produced by the agent includes structured justification, referencing the triggering assessments, linked policy status (or absence), and a natural-language summary of the reasoning path.

To demonstrate the feasibility of the proposed framework, consider a simulated use case based on the heritage site of Delos Island, Greece, which is one of the ARGUS project pilots. The selected threat type is Extreme Wind, known to affect exposed island environments and referenced in preliminary threat identification phases of the ARGUS deliverables.

Table 1. Simulated ABC Risk Assessments for Delos Island under ‘Extreme Wind’ (2012–2023).

Year	A (Frequency)	B (Loss)	C (Fraction)	MR
2012	1	1	1	3
2013	2	1	2	5
2014	3	2	2	7
2015	2	3	2	7
2016	3	3	3	9
2017	3	4	3	10
2018	4	4	4	12
2019	3	5	4	12
2020	4	5	5	14
2021	3	5	5	13
2022	4	5	5	14
2023	5	5	5	15

presents a simulated sequence of ABC risk assessments applied to the site over the period 2012–2023. The values represent a plausible progression of the three ABC components, loosely calibrated to reflect the escalating exposure of the Delos site to wind-driven degradation, informed by qualitative observations in the ARGUS project. The trajectory of $MR$ over time exhibits a monotonic linear increase, reflecting realistic risk evolution under gradual environmental deterioration, punctuated by occasional accelerations in impact or frequency.

The simulated risk escalation reflects a realistic trajectory under unmanaged conditions. The frequency component (A) fluctuates between 1 and 5, capturing the stochastic yet intensifying pattern of extreme wind events, potentially driven by climate change. In contrast, the loss per item (B) and the fraction of the asset affected (C) exhibit gradual and monotonic increases, modeling the cumulative nature of structural deterioration and expanding exposure. This deterministic growth mirrors real-world scenarios where, in the absence of intervention, degradation processes compound over time. The resulting non-linear increase in the total risk score ($MR$) effectively simulates how heritage assets can transition from tolerable to critical risk states, thereby validating the need for temporal reasoning and policy activation by memory-enabled agents.

An agent equipped with the proposed semantic framework queries the knowledge base and detects a steadily increasing Magnitude of Risk ($MR$) trajectory. The absence of an associated Policy linked to the evolving ‘Extreme Wind’ threat triggers the agent’s reasoning mechanisms, based on embedded ABC thresholds and temporal pattern recognition. A justification is generated as follows:

Delos Island has experienced a steady increase in Magnitude of Risk under the ‘Extreme Wind’ threat, from $MR=4$ (2012) to $MR=15$ (2023). No preservation policy is currently active for this threat. Based on internal thresholds and ABC trend cognition, the site is flagged for urgent intervention.

This walkthrough demonstrates how agentic memory, structured around ABC components, enables semantic agents to trace longitudinal risk trends and support explainable prioritization decisions in the absence of direct intervention policies. The example illustrates how the proposed architecture enables coherent, value-sensitive, and traceable reasoning aligned with cultural preservation priorities. In our use case, each Policy instance may encapsulate a specific ABC-based RiskAssessment configuration, allowing the DecisionContext to trace how quantitative risk estimates influence selected decisions.

To ensure interpretability and human oversight, each agentic output is coupled with a provenance chain: documenting the data sources (e.g., sensor input or manual assessment), the $MR$ score evolution, the applicable rule activation, and the associated asset state. These outputs can be stored semantically and may also be rendered into curated explanations accessible via human-readable interfaces. This capability opens the door to future applications in risk storytelling, where an agent may narrate the evolving vulnerabilities of an asset across time, grounded in documented observations and formalized risk logic.

4. Proof-of-Concept Simulation of Risk Reasoning with Agentic Memory

To assess the operational validity of the proposed framework, we implemented simulated Magnitude of Risk ($MR$) trajectories (A Colab notebook implementing the ABC risk simulations and visualizations for selected pilot cases is available at: https://colab.research.google.com/drive/12BYY2K7aiLop8cTRPwonbn98x3y8pOaF, (accessed on 3 June 2025)) for three heritage pilot assets from the ARGUS project: Delos Island (Greece), the Cellar Town of Baltanás (Spain), and the Abbey of Sant’Antonio di Ranverso (Italy). For each site, a principal environmental threat was modeled based on context-specific vulnerabilities identified in project deliverables. Two comparative scenarios were constructed over a 12-year period (2012–2023): (a) a baseline with no preservation policy activation and (b) an intervention case with a policy introduced at some point in time, different per pilot. In this proof-of-concept, agentic behavior is illustrated through predefined procedural rules operating over time-indexed risk memory. Simulated agents identify critical thresholds, detect policy gaps, and generate recommendations based on the trajectory of the computed $MR$. While policy generation is not automated, the reasoning mechanism autonomously flags intervention needs and, in a full semantic implementation, would justify them through linked ontological traces.

Delos Island faces increasing weathering pressures due to salt crystallization, which is exacerbated by wind and rain during winter months. These environmental stressors progressively weaken exposed masonry and porous materials, leading to surface loss and scaling. In the simulation, salt weathering is modeled as the principal threat, with ABC components increasing over time. A hypothetical policy (e.g., hydrophobic treatment or seasonal shielding) is modeled as a partial containment strategy that stabilizes the progression of risk without full elimination.

In Baltanás, the dominant threat arises from humidity accumulation, which—compounded by heavy rainfall—has contributed to structural failures in the historical cellar network. The simulation targets this primary hazard, reflecting a scenario in which prolonged exposure increases both the severity and spatial spread of degradation. A policy involving environmental stabilization (such as drainage systems or active dehumidification) is introduced in the model from 2020 onward, yielding a noticeable deceleration in the ABC rise and illustrating mitigation effects.

The Ranverso site is affected by several stressors, including thermal fatigue and possible micro-vibrations from traffic or geological shifts. For clarity and alignment with the single-threat simulation, we model thermal fatigue as the representative threat. This hazard induces gradual material strain and expansion cracks over time. The simulated $MR$ shows moderate but persistent growth under unmanaged conditions. A targeted policy scenario, combining environmental monitoring and selective reinforcement, is introduced from 2021, leading to a slower rate of ABC progression while deterioration continues.

The simulation data for each pilot site were generated to reflect plausible ABC dynamics consistent with the nature of the threat, as documented in ARGUS project pilot analysis. For each case, a set of $MR$ trajectories was computed over a 12-year period, including a scenario without mitigation and one with policy intervention. ABC components were generated to follow plausible progressions, in which frequency (A) fluctuates randomly, while severity (B) and exposure (C) increase gradually over time. Upon policy activation, A, B, and C values continue evolving with localized noise, but the computed $MR$ is reduced by a fixed offset to simulate mitigation effects. This yields diverging $MR$ trajectories while preserving the same pre-policy history across both scenarios. The numeric generation logic underlying these scenarios follows simple procedural rules. The frequency component (A) is generated using random integer sampling in the range $[1,5]$, representing year-to-year stochastic variation in threat recurrence. The loss per item (B) and fraction of the asset affected (C) increase deterministically over time, implemented as stepwise functions: B increases approximately every three years, and C every four years. These increments simulate gradual escalation in damage severity and spatial extent. For the policy-intervention scenario, mitigation is implemented by subtracting a fixed offset (typically $\Delta MR=3$) from the computed Magnitude of Risk ($MR=A+B+C$) in all years following policy activation, while keeping the underlying A, B, and C values intact. This ensures traceability and allows semantic agents to assess the temporal divergence between unmanaged and managed risk. No statistical model or real-world dataset is employed; instead, the simulation constructs representative risk trajectories designed to test the framework’s ability to support structured agentic reasoning.

The simulated interventions differ by case purely for illustrative purposes: in Delos, the policy models wind shielding that slows down escalation post-2018; in Baltanás, humidity control introduced in 2020 yields a similar containment effect; in Ranverso, introduced measures produce only partial suppression, consistent with structural risk being harder to reverse. All simulations strictly adhere to the ABC risk framework, with A, B, and C assigned integer values between 1 and 5. This constraint ensures that computed $MR$ values remain in the valid range [3, 15], enabling realistic agentic reasoning. The modeling logic mirrors cognitive processes by encoding:

• Episodic memory: each yearly observation is stored as a distinct RiskAssessment instance, timestamped and annotated with discrete ABC values.
• Semantic memory: structured ontological links associate each RiskAssessment with the corresponding Asset, Threat, and, if applicable, an active Policy.
• Policy-aware reasoning: agents continuously monitor $MR$ trends per heritage asset. In the absence of a Policy, a significant or accelerating increase in $MR$, relative to the asset’s historical pattern, triggers internal prioritization. Once a Policy is active, the agent evaluates post-intervention $MR$ divergence to assess mitigation impact and support further decision-making.

This structure enables the simulation of a minimal yet operational agentic AI loop: memory is not passive storage but actively interpreted through semantic reasoning. While no deterministic prediction model governs the $MR$ series, the scenarios encode plausible environmental dynamics and intervention effects. Figure 1 presents the results. The visualizations underscore how semantically structured memory and per-site policy logic support explainable, context-aware monitoring in heritage-focused AI agents.

Several limitations should be acknowledged in this simulation. First, the use of synthetic ABC values and modeled policy effects, which may not capture all real-world complexities. Second, the reasoning engine currently operates synchronously and serially, limiting responsiveness in larger deployments. Third, while semantic explainability is preserved, integration into operational heritage workflows has not yet been empirically validated. Addressing these limitations requires real-world piloting, stakeholder co-design, and extensive testing of reasoning performance at scale.

In practice, the ABC method depends on expert judgment, which can be subjective or inconsistent across assessors. This introduces uncertainty into the A, B, and C components that, in real-world deployments, would require additional handling. Future agentic systems could incorporate probabilistic reasoning, fuzzy logic, or consensus mechanisms to accommodate ambiguity and conflicting inputs.

Moreover, policy effects in our simulation were implemented as fixed $MR$ offsets, purely for illustrative purposes. Actual preservation interventions often have delayed, nonlinear, and context-dependent impacts, which may require more sophisticated modeling approaches, such as causal inference frameworks or scenario-based simulation.

These aspects, while important for full-scale implementation, fall outside the scope of this study. The present work aims to establish the foundational framework for memory-enabled, semantically grounded agentic AI in heritage preservation. Detailed modeling of expert disagreement, intervention dynamics, or optimization is left for future research, to be pursued once empirical deployments are underway. The proposed approach, however, is intentionally designed for modularity and semantic scalability and could ultimately support distributed, policy-aware reasoning over broader geographies, even at national or continental scales.

5. Discussion

The integration of structured risk memory into agentic AI architectures presents a shift in cultural heritage management, from reactive monitoring toward anticipatory reasoning grounded in temporal and semantic context. By embedding time-stamped ABC risk assessments within the Panoptes ontology, agents can track evolving threats, evaluate policy impacts, and provide traceable justifications for conservation recommendations. Risk is thus modeled not as a static value but as a trajectory, enabling the early detection of escalating vulnerabilities.

As demonstrated in Section 4, the agentic framework can operate on semantically structured, threat-specific inputs to assess policy effectiveness, detect threshold crossings, and compare counterfactual outcomes. The Delos, Baltanás, and Ranverso scenarios illustrate how discrete, integer-valued ABC components can encode plausible preservation dynamics, while preserving interpretability. Through simulated $MR$ trajectories, agents expose differences in unmanaged versus managed risk profiles and enable risk flagging based on formally represented knowledge.

The presented simulation treats policies as predefined and manually linked to specific threat contexts. While, in this setting, agents do not learn or generate policies autonomously, they reason over their presence, absence, and temporal alignment by querying structured memory, specifically by evaluating whether or not a Policy instance exists for a given Threat and Asset, and comparing its activation date against the temporal trajectory of risk escalation encoded in successive RiskAssessment instances. In our simulation, this reasoning is externally visible through divergences in $MR$ curves, policy-aware alerts, and threshold-based evaluations across time. Although limited, yet transparent, this deterministic policy representation demonstrates the system’s explainability and points toward future extensions, in which involving agents could adapt or learn policy inference grounded in evolving risk patterns.

At the same time, introducing memory-enabled AI into cultural heritage brings specific responsibilities. The underlying ontology must clearly represent what matters most: what is considered valuable, what is at risk, and what deserves priority. These definitions should not be decided by technologists alone. If the system is developed without input from heritage professionals or local communities, it may overlook important cultural differences or reinforce unintended biases. For this reason, cultural values must be integrated into the design of the ontology from the beginning, not added later as an afterthought. This is mostly supported by the Policy instances.

Technically, while the use of SPARQL-based reasoning supports modular and explainable decision logic, it may pose challenges in scenarios requiring high-frequency updates or multi-site monitoring. Future extensions may explore hybrid semantic-procedural reasoning models, asynchronous policy propagation mechanisms, or integration with edge-computing agents to ensure scalability. A minimal toy example of agent-based reasoning using SPARQL is shown in Listing 2.

Listing 2. SPARQL query to detect risks that fall outside policy coverage intervals.

PREFIX pan: <http://www.argus-project.eu/panoptes#>

PREFIX time: <http://www.w3.org/2006/time#>

SELECT ?asset ?risk ?policy

WHERE {

?risk a pan:RiskAssessment ;

pan:MagnitudeOfRisk ?mr ;

pan:hasTime ?t ;

pan:assesses ?threat ;

pan:ofAsset ?asset .

?policy a pan:Policy ;

pan:relatedTo ?threat ;

pan:hasTime ?interval .

?interval time:hasBeginning ?start ;

time:hasEnd ?end .

FILTER NOT EXISTS {

?t time:inXSDDateTimeStamp ?timeVal .

FILTER (?timeVal >= ?start && ?timeVal <= ?end)

}

}

Another promising dimension is the narrative generation of evolving $MR$ patterns into human-readable alerts. Such functionality could enhance curatorial workflows by aligning AI outputs with the interpretive practices of preservation and conservation experts.

In sum, this work contributes a conceptual and semantic foundation for Cultural Agentic AI: systems that remember, reason, and respond, not to replace human judgment but to augment it through foresight, transparency, and alignment with cultural values.

It is important to reiterate that the simulations presented in this work are intended as conceptual illustrations of the proposed framework’s capabilities, not as empirical validations. The generated data reflect plausible but synthetic risk trajectories and policy effects, designed to test whether or not agentic reasoning mechanisms can operate over structured memory in a transparent and semantically grounded way. Future work must engage with real-world ABC data from field assessments and calibrate the system under operational constraints.

A related challenge lies in the translation of real-world conservation policies—often narrative, qualitative, or multi-objective—into formal, machine-readable constructs such as pan:Policy instances. Capturing trade-offs, resource constraints, and context-specific value priorities will require richer ontological structures and participatory encoding approaches. While our present model uses predefined policies as semantic anchors, a fully operational system will need to bridge this abstraction gap in collaboration with domain experts.

6. Conclusions and Future Work

This work proposed a semantic and theoretical foundation for agentic AI in cultural heritage by embedding structured risk memory and policy-aware reasoning within the previously published Panoptes ontology. By integrating the ICCROM/CCI ABC model, the framework enables agents to assess evolving threats, store temporally grounded knowledge, align actions with preservation policies, and offer justifications for decision-making.

Using simulated pilot cases from the ARGUS project, we demonstrated how such agents can monitor risk trajectories, identify threshold crossings, and evaluate the counterfactual effects of mitigation strategies. These proof-of-concept scenarios underscore the feasibility of memory-guided, semantically explainable reasoning as a foundation for anticipatory conservation planning. This shift, from reactive monitoring to goal-directed, memory-enabled inference, positions AI not merely as a sensor integrator but also as a cognitive collaborator in the preservation of heritage assets. While the presented simulation framework operates on predefined and manually encoded policies, the underlying architecture is extensible toward agents that autonomously infer or adapt policy logic based on accumulated risk memory. While these simulations do not constitute empirical validation, they serve as demonstrations of conceptual feasibility. Future work must validate the framework on real data from ARGUS pilot sites.

The result is a semantically structured, culturally aligned agentic architecture, one that does not merely monitor, but remembers, reasons, and advises in defense of our shared heritage.

Future work directions include:

• Implementing SPARQL-based reasoning pipelines and operational agents.
• Deploying and validating the system in real-world ARGUS pilot sites.
• Integrating natural language generation to narrate agentic decisions and alerts.
• Evaluating explainability, stakeholder alignment, and cultural fit through participatory testing.