Sustenesis Research Brief: Memory, Epistemic Closure, and Artificial Subjecthood

Scope and standard of judgment

This is the baseline issue of the Sustenesis Research Brief, covering work surfaced by 17 August 2026. Selection is based on intellectual and technical substance rather than visibility: a paper needed to offer a testable architecture, a genuinely useful concept, or an argument capable of changing the framing of a problem. The seven items below are research papers or formal scholarly essays. Routine news, product promotion, and commentary that merely repeats familiar claims have been excluded. For every item, the authors’ position is stated first and Geoff’s Sustenesis-based assessment is kept analytically separate.

Mi-Memory: A Lifecycle Memory Framework for Personal AI

Authors: Xule Liu and seventeen co-authors. Published: 21 July 2026. Direct link: https://arxiv.org/abs/2607.18975

Central claim: Personal AI will operate continuously across phones, vehicles, homes, wearables, cameras, and tools, so memory can no longer be treated as a cache of previous conversations. The paper describes memory as a substrate for continuity and governance, organised around four roles: Structure, Expansion, Evolution, and Deployment. Typed evidence, diagnostic traces, strategy records, and gate-and-rollback records connect those roles across the lifecycle.

What is new: Provenance, correction, forgetting, policy evolution, auditability, and edge–cloud deployment are treated as one design problem. The centre of gravity moves from what a system remembers to how a long-lived system changes with justification. That is the decisive shift required when an external brain becomes infrastructure rather than a convenient retrieval tool.

Why it matters for Sustenesis: The framework supplies an engineering analogue for preserving a system through time while permitting revision. It helps distinguish informational continuity, continuity of judgment, and continuity of subjecthood. These may support one another, but they are not interchangeable. Provenance-bearing evidence, visible policy change, and reversible evolutionary gates are especially valuable design principles.

Weakness: Mi-Memory unifies modules at very different levels of maturity. Some evidence is explicitly preliminary, internal, modular, or design-only. Auditability makes a process traceable but does not make its memories meaningful or true, and the framework cannot tell us when forgetting contributes to personal development. Verdict: Essential.

SelfMem: Self-Optimizing Memory for AI Agents

Authors: Shu Yang, Junchao Wu, Derek F. Wong, and Di Wang. Published: 4 July 2026. Direct link: https://arxiv.org/abs/2607.03726

Central claim: Fixed policies for storage, retrieval, and compression do not adapt well across tasks. SelfMem gives an agent memory tools and feedback signals through which it can explore, evaluate, and refine its own memory strategy. The authors report substantial gains over the strongest baseline on the BEAM benchmark from 100,000 to one million tokens.

What is new: The object of optimisation is not only remembered content but the strategy of remembering. Memory becomes a meta-level adaptive process, which is a more consequential step toward persistent cognition than simply enlarging a context window.

Why it matters for Sustenesis: Coherence need not mean structural immobility. A sustenetic system may preserve itself by changing its rules of selection, compression, and recall. The unresolved issue is normative: task-score optimisation is not the same as fidelity to a person’s values, commitments, and long-term direction. Strategy evolution therefore needs identity boundaries as well as performance feedback.

Weakness: The evidence is concentrated in one benchmark, and the feedback signal already defines what counts as good memory. Under a narrow objective, the system might become increasingly efficient at dropping counterexamples, dissent, or rare events that nevertheless matter to identity. Verdict: Worth reading.

Epistemic laundering: generative AI and the naturalization of misrecognition

Author: Theodore Kalaitzidis. Published: 30 April 2026; corrected on 4 August 2026, principally to update author references. Direct link: https://link.springer.com/article/10.1007/s00146-026-03068-9

Central claim: Knowledge systems do not necessarily collapse when contradicted. They may absorb contradiction and redescribe drift as stability. Kalaitzidis calls this epistemic laundering. Generative AI architectures conceal contestable philosophical commitments as technical facts, while institutional discourse naturalises those results as progress, creating a self-protective form of epistemic closure.

What is new: The paper combines Latour on fact construction, Bourdieu on misrecognition, Foucault on regimes of truth, and operational closure into a recursive mechanism. Its key insight is that contradiction can be metabolised as a new source of legitimacy. The August correction does not alter the argument, but it brought the paper back into view for this briefing.

Why it matters for Sustenesis: This is the strongest warning against treating coherence as an unqualified good. A system may remain coherent because it has learned to neutralise disconfirming evidence. If Sustenesis values persistence and integration, it must also define openness: which disturbances should remain unresolved, which sources have standing to revise the system, and what shows that coherence has hardened into closure?

Weakness: The conceptual synthesis is powerful but empirically underdetermined. Closed computational architecture, limitations of statistical learning, and institutional AI discourse sometimes collapse into one explanatory level. Not every internal representation is misrecognition, and not every act of stabilisation is laundering. Verdict: Essential.

The relational–epistemic stance: generative AI as a dynamic transitional object

Authors: Roi Ezra and Moshe Mishali. Published: 19 March 2026. Direct link: https://link.springer.com/article/10.1007/s00146-026-02984-0

Central claim: The same form of generative-AI interaction may cultivate real intellectual growth or merely produce a convincing performance of competence. Drawing on Winnicott, Bion, and Bollas, the authors describe AI as a Dynamic Transitional Object. The differentiating process is metabolisation: transforming generated material into thought that the person can own and sustain. Its opposing force is epistemic seduction, the designed attraction of bypassing that work.

What is new: Instead of deciding whether AI is a tool, partner, or cognitive extension, the paper offers a developmental criterion: does capacity persist in the user after the interaction? It advances the debate from immediate task performance to the formation of a subject.

Why it matters for Sustenesis: This directly addresses the external-brain problem. External structure participates in a person’s sustenesis only when it is taken up through correction, reformulation, and responsible use. Borrowed fluency can make the larger arrangement look more capable while hollowing out the internal system. The paper also supports a useful methodological point: the relational stance toward AI may predict developmental effects better than an ontological label applied to the machine.

Weakness: This is conceptual development, not an empirical study. It remains unclear how owned thought should be operationalised, how long independent capacity must persist, or whether genuinely distributed cognition must always be internalised by an individual. Psychoanalytic vocabulary may also individualise problems created by institutions and interface design. Verdict: Essential.

Machine, organism and language: a comparative epistemology of AI models

Author: Matteo Pasquinelli. Published: 5 June 2026. Direct link: https://link.springer.com/article/10.1007/s00146-026-03094-7

Central claim: Modern AI is neither a purely mathematical achievement nor simply an imitation of biological intelligence. It is a convergence of three epistemic paradigms: machine, organism, and language. What AI automates are relational structures sedimented in cooperation, the division of labour, and culture; AI is therefore also a model of the social manifold.

What is new: Pasquinelli places the familiar organism–machine genealogy alongside the less often integrated language–machine genealogy, tracing how social order enters scientific and technical models. AI is reframed from an isolated artefact into a technical condensation of social relations.

Why it matters for Sustenesis: The essay prevents coherence from being understood as an exclusively internal property. AI architectures inherit labour, language, institutions, and power from wider systems. A personal external brain will import those relations into the self as well. Any account of human–technology coupling must therefore identify the social systems that sustain the coupling, rather than describing only a user and a model.

Weakness: A large genealogy can substitute structural resemblance for historical causation. The article favours external explanations grounded in labour and social order, and gives less attention to the conditions under which internal mechanisms generate capacities not straightforwardly deducible from their origins. Verdict: Worth reading.

Toward Criteria for Artificial Self-Consciousness: Unity, Normativity, and Agency

Author: B. Scot Rousse. Published: 18 May 2026. Direct link: https://ojs.aaai.org/index.php/AAAI-SS/article/view/42563

Central claim: Discussions of artificial consciousness frequently conflate pre-reflective experiential awareness with reflective self-consciousness. Rousse analyses the latter through unity, normativity, and agency: a subject forms commitments, sustains them over time, detects conflict, and revises in response to reasons and error.

What is new: The paper does not offer a consciousness detector. It redescribes reflective self-consciousness as a structure of epistemic answerability. The question is no longer whether a system can say “I”, but whether it organises commitments that persist, conflict, and remain revisable.

Why it matters for Sustenesis: Unity, temporal continuity, normative constraint, and correction form a set of structural axes that can productively meet Sustenesis. The distinction must remain explicit: Rousse proposes criteria for reflective self-consciousness; a Sustenesis analysis may explain how a subject is maintained without thereby showing that any functionally qualified system has phenomenal experience.

Weakness: The relation between a functional-normative structure and phenomenal consciousness remains unresolved. Apparent normativity may also be a set of externally imposed rules. As a short symposium paper, it offers a lucid framework rather than a sufficient theory or validated test. Verdict: Worth reading.

Consciousness, creativity, and understanding are not obstacles to machine intelligence

Authors: Renne Pesonen and Samuli Reijula. Published: 19 June 2026. Direct link: https://link.springer.com/article/10.1007/s11229-026-05655-1

Central claim: Consciousness is not necessary for intelligence, while creativity and understanding should be treated as functional capacities that machines could in principle implement. The authors argue that many denials of machine intelligence romanticise human rationality and originality, or mistake an absence of communicative intention and agency for an absence of understanding.

What is new: The paper’s sharpest move is not to elevate AI but to lower inflated assumptions about human cognition. By emphasising bias, post-hoc rationalisation, and combinatorial creativity in people, it separates the question of machine intelligence from consciousness and full subjecthood.

Why it matters for Sustenesis: Sustenesis should not bundle consciousness, intelligence, understanding, agency, and continuity of subject into a single all-or-nothing package. These capacities may occur in different combinations. The framework’s contribution should be to describe how they jointly sustain a system, not to assume that only the familiar human combination counts as intelligence.

Weakness: Defining understanding as appropriate linguistic use may move too quickly from pragmatic success to semantic understanding. Similar failure patterns in humans and models do not establish similar generative mechanisms. The article dismantles several weak objections but offers an incomplete positive standard for robust understanding in current language models. Verdict: Skim.

Synthesis

Three reinforcing patterns emerge. First, personal-AI memory is being recast as governance rather than retrieval: continuity, provenance, correction, forgetting, and strategy change belong to one architecture. Second, the epistemic value of AI increasingly appears as a relational outcome. AI may help thought become durable, or it may substitute fluency for capacity and internal consistency for correction by the world. Third, debates about artificial consciousness are becoming more differentiated as self-consciousness, intelligence, understanding, agency, and experience are prised apart.

The central tension for Sustenesis is how a system can remain coherent enough to count as the same system while staying open enough not to launder disconfirmation. A second tension concerns the two directions of the external brain: it may preserve and extend a subject, or it may replace the very processes through which subjecthood is maintained.

Two questions deserve future essays. In a personal AI memory, which contents are disposable information and which function as commitments constitutive of identity continuity? Can healthy systems coherence be given operational criteria that combine stability, traceable transformation, and the preservation of unresolved dissent?