Sustenesis Research Brief: Memory Gates, Cognitive Friction, and Normative Interfaces

About this issue

This is the second Sustenesis Research Brief. The baseline issue, completed on 17 August 2026, covered seven works on lifecycle memory for personal AI, self-optimising memory, epistemic laundering, AI as a dynamic transitional object, comparative epistemologies of AI, and artificial self-consciousness. None is repeated here. This issue instead examines seven strong papers newly surfaced in the subsequent search. Their shared concern is not how much an AI system can store, but what should be admitted into it, how experience becomes knowledge, how external support can preserve rather than replace judgment, and where responsibility should be stabilised once AI participates in action.

Beyond Similarity: Trustworthy Memory Search for Personal AI Agents

Authors: Jiawen Zhang, Kejia Chen, Jiachen Ma, Yangfan Hu, Lipeng He, Yechao Zhang, Jian Liu, Xiaohu Yang, Tianwei Zhang, and Ruoxi Jia. Published: 4 June 2026. Direct link: https://arxiv.org/abs/2606.06054

Central claim: Long-term memory in personal AI cannot be governed by semantic similarity alone. A memory may resemble the present query while belonging to the wrong domain, identity, or context, creating cross-domain leakage, sycophancy, tool-call drift, or memory-induced jailbreaks. The authors define memory search as a trust boundary and introduce MemGate, a lightweight module that decides whether candidate memories should enter the model context.

What is genuinely new: Memory is treated as a durable control channel rather than a passive information layer. The intervention sits at admission time and can be deployed without changing the backbone model or rewriting the memory store.

Why it matters for Sustenesis: Continuity does not require carrying the whole past into every present situation. Statistical relevance is only proximity among differences; sustenesis also requires boundaries, context, and identity constraints. A personal external brain therefore needs authority over memory admission, not merely a larger capacity to remember.

Weakness: MemGate still inherits a training objective that defines appropriateness. It can reduce specified threats but cannot determine the significance of a memory within a life, or settle who may change its admission rules. Verdict: Essential.

Hypothesis-Driven Reasoning for Large Language Models

Authors: Aakash Kumar Agarwal and Moyuru Yamada. Published: 14 March 2026. Direct link: https://ojs.aaai.org/index.php/AAAI/article/view/37146

Central claim: Even when an LLM receives enough multimodal experience, it may fail on a new task because it cannot extract reusable structure from noisy episodes. Hypothesis-Driven Reasoning first identifies potential factors and then builds explicit semantic memory through a generate-and-verify loop. Oracle hypotheses lifted accuracy on the authors’ new task from 35.3 to 92 per cent, while the implemented method approached that upper bound.

What is genuinely new: The method does not add another retrieval pass. It compresses episodes into hypotheses that can be stated, checked, and revised. Memory begins to function as a knowledge structure rather than an event archive.

Why it matters for Sustenesis: This closely matches the view that information becomes knowledge when it can be preserved, retrieved, tested, corrected, and made to work reliably. Explicit hypotheses form an intermediate layer between differences in experience and a sustained structure. An external brain should not merely preserve a person’s past; it should help form inspectable conceptual relations.

Weakness: The experiment relies on a task created by the authors, and the dramatic oracle result does not establish comparable robustness in an open world. Explicit hypotheses may also close interpretation prematurely by forcing poorly understood experience into the wrong structure. Verdict: Worth reading.

Sensorimotor Regularities as Alignment between Humans and Large Language Models

Authors: Jingyi Li, Jinghui Hu, and Per Ola Kristensson. Published: 15 April 2026. Direct link: https://doi.org/10.1145/3778353

Central claim: Using image schemas from cognitive linguistics, the authors compare human and LLM conceptual representations. Three models reproduce many human sensorimotor schemas, but divergence grows when schemas are mapped to concepts and when several schemas must be combined. Supplying targeted sensorimotor priors made outputs clearer, more context-sensitive, and more human-like in a downstream evaluation.

What is genuinely new: Embodiment is operationalised rather than invoked as a slogan. The paper distinguishes schema distribution, schema-to-concept association, and schema co-occurrence. The resulting gradient is revealing: models resemble humans at the level of available components but less so in relational composition.

Why it matters for Sustenesis: Similar outputs do not establish similar generative structures. Language can preserve relational traces of bodily experience, enabling a disembodied model to reproduce part of that structure. Reproduction is not possession of the originating experience. Structural inheritance, functional similarity, and subjective experience must remain separate questions.

Weakness: The human sample contains only 24 participants, the downstream study 45, the conceptual domain is narrow, and the models are no longer current. Image-schema theory is also only one account of conceptual structure. Verdict: Essential.

Scaffolding Critical Thinking with Generative AI: Design Principles for Integrating Large Language Models in Higher Education

Authors: Mireia Vendrell and Samantha-Kaye Johnston. Available online: 7 March 2026; published in the June 2026 volume. Direct link: https://doi.org/10.1016/j.caeai.2026.100572

Central claim: Unstructured use of LLMs can produce cognitive offloading, metacognitive disengagement, and reduced epistemic agency. The authors derive eight design principles from cognitive psychology, educational theory, and AI ethics. The most important preserve cognitive friction, treat AI as a provisional thinking partner, embed evaluation throughout learning, and alternate AI-mediated with AI-free phases.

What is genuinely new: The paper moves beyond the sterile choice between permitting and banning AI. It treats interaction sequence as the design problem and cognitive friction as a condition of judgment rather than an efficiency defect.

Why it matters for Sustenesis: If an external system organises, judges, and expresses before difficulty reaches the person, it can sustain output while weakening the internal structures that produce it. Sustenesis must distinguish the continued completion of tasks from the continued formation of capability. These trajectories can move in opposite directions.

Weakness: The framework is conceptual and normative, and its eight principles lack longitudinal controlled evidence. AI-free phases in education also cannot be transferred without qualification to every professional workflow. Verdict: Essential.

Agency and alignment: toward a normative architecture for human–AI interaction

Authors: Saša Josifović and Jörg Noller. Published: 11 April 2026. Direct link: https://link.springer.com/article/10.1007/s00146-026-02950-w

Central claim: Alignment should not be understood as inferring a fixed set of human values from data. It should mean structurally integrating AI into existing normative domains. The authors connect extended human agency, practical autonomy, and a normative interface. AI may extend human purposes without becoming an autonomous moral subject, but its outputs must enter institutions where reasons can be requested, decisions contested, and responsibility allocated.

What is genuinely new: Alignment moves from internal value learning to relational and institutional architecture. A normative interface is not an ethical prompt but an action space connecting purposes, reasons, roles, appeal, and accountability.

Why it matters for Sustenesis: Systems coherence cannot be measured entirely within a model. It also depends on whether the relations among model, person, institution, and responsibility can be maintained. Norms are structural constraints on how a system continues, changes, and assigns consequences.

Weakness: The argument is relatively optimistic about the normative capacity of law and institutions. Institutions can themselves become rigid, exclusionary, and protective of established power. A normative interface may stabilise an unjust system if it merely conforms AI to existing rules. Verdict: Essential.

Can generative artificial intelligence be considered a cognitive subject? An analytic analysis

Authors: Steven S. Gouveia and Yinchun Wang. Published: 19 February 2026. Direct link: https://link.springer.com/article/10.1007/s00146-026-02924-y

Central claim: Generative AI may be a cognitively significant contributor to knowledge production without being a cognitive subject. The authors propose five conditions: information competence, reasons-responsiveness, metacognitive self-representation, capacity for consciousness, and robust intentionality. Current systems satisfy only the weaker part of this set.

What is genuinely new: The paper cleanly separates cognitive contribution from cognitive subjecthood. It also develops the idea of third-order knowledge, a hybrid product of training data, model reconstruction, user prompting, selection, and editing. This recognises AI’s substantive epistemic role without automatically assigning it subjecthood and responsibility.

Why it matters for Sustenesis: The paper directly intersects Geoff’s question of whether knowledge requires a knowing subject, but the two issues should not be collapsed. The authors examine qualification as a cognitive subject. Sustenesis can ask separately whether a stable, usable, testable, and revisable knowledge structure must remain owned by a single continuing subject. The claim that AI is not a subject does not by itself show that an AI system contains no knowledge.

Weakness: The five conditions are presented as jointly necessary and sufficient, although consciousness and robust intentionality remain disputed concepts. A demanding definition can predetermine the negative result and may understate degrees of subjecthood in distributed epistemic systems. Verdict: Worth reading.

AI and Semantic Pareidolia: When We See Intelligent Consciousness where there is None

Author: Luciano Floridi. Published: 19 February 2026. Direct link: https://link.springer.com/article/10.1007/s13347-026-01052-1

Central claim: People readily attribute meaning, intention, emotion, and consciousness to systems that do not possess them. Floridi calls this semantic pareidolia, by analogy with seeing faces in clouds. Digital immersion, social isolation, commercial incentives, and technical progress amplify the tendency, allowing ordinary anthropomorphism to become technological idolatry.

What is genuinely new: The main contribution is conceptual focus and a memorable name. Evidence for AI consciousness is shaped not only by model behaviour but by user psychology, interface design, and commercial narrative.

Why it matters for Sustenesis: This adds the observer’s side to subject formation. Whether a system is treated as a subject depends partly on sustained relational projection, not only on internal organisation. Yet Sustenesis should not reduce consciousness entirely to projection. Explaining attribution does not prove that the attributed system necessarily lacks experience.

Weakness: This is a short editor’s letter with limited argument and empirical support, and its title assumes the absence of consciousness in advance. It identifies a genuine epistemic risk but does not decide the ontological status of AI experience. Verdict: Skim.

Synthesis

The clearest shift in this issue is from asking what a system contains to asking what may enter it and what constrains the result. MemGate regulates memory admission. Hypothesis-driven reasoning turns episodes into testable structure. Sensorimotor analysis separates linguistic appearance from generative organisation. Cognitive friction marks the boundary between external support and internal capability. The normative interface then connects model behaviour to reasons, institutions, and responsibility.

This places two pressures on Sustenesis. First, sustained coherence must include principled refusal as well as preservation and integration. What a system excludes in order to remain itself matters as much as what it retains. Second, an external brain cannot be judged only by immediate output. If it keeps work flowing while the person’s structures of judgment recede, it sustains a human–machine workflow but may not sustain the human subject.

Two questions deserve development. Who should hold memory-admission authority in personal AI: the user, the model, an institution, or a layered system in which each can constrain the others? When a human–AI arrangement reliably produces valid knowledge although no single participant fully understands the production process, how should knowledge, subject responsibility, and system responsibility be separated?