What Was I Thinking? A Framework and Survey of Metacognitive Experience Memory in Cognitive Architectures
Veröffentlichungsdatum
2026-09-08
Betreuer
Gutachter
Zusammenfassung
For many modern Artificial Intelligence (AI) systems, it is nearly impossible to understand how specific outputs arise from given inputs. This includes Large Language Models (LLMs), which can generate plausible-sounding explanations of their own processes; initial case studies, however, cast doubt on the faithfulness of these explanations, suggesting that they do not necessarily reflect the underlying processing that actually produced the output. Even post-hoc Explainable Artificial Intelligence (XAI) methods usually show only which inputs were influential, not how a model used them to arrive at a result. This lack of transparency gives rise to a range of problems, including severe limitations in epistemological insight, security guarantees, calibrated user trust, human oversight, debuggability, and meaningful autonomy. In contrast, human cognition is transparent in that it is aware of its own processes through metacognitive experiences and can usually report them reliably. From a biomimetic perspective, this raises the question of how AI systems might exploit metacognitive experiences, not least in order to increase their transparency. Research on metacognition in AI, however, has so far remained a highly fragmented field: divergent theories, terminologies, and design principles have led to disjointed developments and severely limited comparability across relevant approaches. A principled account of metacognitive experiences is still very much missing.
This cumulative dissertation addresses that gap by examining the metacognitive experiences of an established class of AI systems specifically designed to model human cognition: Computational Cognitive Architectures (CCAs). It does so in three steps: first, a systematic review synthesises three decades of research on how existing CCAs model metacognitive experiences, store them in episodic memory, and later utilise them. Second, drawing on these principles, this thesis proposes a standardised vocabulary and representational schema for metacognitive experiences in the form of the operational ontology MOI. Third, it explores how CCAs could reason about such representations — which requires both database queries and characterisations of abnormal cognition — by establishing the decidability and complexity of a range of expressive description logics with circumscription.
Taken together, the three methodologically diverse studies contribute a coherent framework for metacognitive experiences in CCAs spanning empirical, representational and computational aspects. Specifically, they provide concrete recommendations on how future CCA research should proceed and how metacognition could be investigated as a potentially emergent property in opaque systems such as LLMs. In this way, the thesis contributes to the development of more transparent AI systems in the future, as well as to the disciplines of logic, cognitive science, and robotics.
This cumulative dissertation addresses that gap by examining the metacognitive experiences of an established class of AI systems specifically designed to model human cognition: Computational Cognitive Architectures (CCAs). It does so in three steps: first, a systematic review synthesises three decades of research on how existing CCAs model metacognitive experiences, store them in episodic memory, and later utilise them. Second, drawing on these principles, this thesis proposes a standardised vocabulary and representational schema for metacognitive experiences in the form of the operational ontology MOI. Third, it explores how CCAs could reason about such representations — which requires both database queries and characterisations of abnormal cognition — by establishing the decidability and complexity of a range of expressive description logics with circumscription.
Taken together, the three methodologically diverse studies contribute a coherent framework for metacognitive experiences in CCAs spanning empirical, representational and computational aspects. Specifically, they provide concrete recommendations on how future CCA research should proceed and how metacognition could be investigated as a potentially emergent property in opaque systems such as LLMs. In this way, the thesis contributes to the development of more transparent AI systems in the future, as well as to the disciplines of logic, cognitive science, and robotics.
Schlagwörter
Artificial Intelligence
;
Metacognition
;
Ontologies
;
Robotics
;
Knowledge Representation
;
Description Logics
Institution
Institute
Dokumenttyp
Dissertation
Sprache
Englisch
Dateien![Vorschaubild]()
Lade...
Name
Dissertation_Nolte_main_fixed.pdf
Size
7,12 MB
Format
Adobe PDF
Checksum
(MD5):cf778187a9b125bfd857ccde53539289