Sustenesis Research Brief: Memory as State, Runtime Contracts, and Agent Organisation

About this issue

This is the third Sustenesis Research Brief. The previous issue covered memory admission, hypothesis-driven knowledge formation, sensorimotor schemas, cognitive friction, normative interfaces, cognitive subjecthood and semantic pareidolia. None of those works is repeated here. The past day has not produced a paper strong enough to redefine the whole field by itself, so this issue follows the brief’s broader rule: work either recently published or newly surfaced in the search since the previous briefing. Six papers not previously included are selected because they create genuine structural pressure on the developing Sustenesis framework.

Towards a Formal Definition of Agent Memory: Basis, Span, Optimality, and the Sequential Memory Problem

Author: Hongyao Tang. Published: 12 August 2026. Direct link: https://arxiv.org/abs/2608.11654

Central claim: The paper attempts to give agent memory a formal object. Memory is treated as a basis and knowledge as the span generated from that basis; query answerability becomes a coverage problem. Capacity limits, noise and continual writing are then brought into the same framework, with long-term memory formation described as a sequential decision problem.

What is genuinely new: Many memory systems are compared only through benchmark scores without a common theoretical object. This work reframes “how much is remembered” as an optimisation among capacity, coverage, precision and future utility, allowing different architectures to be compared on a shared conceptual axis.

Why it matters for Sustenesis: This directly intersects the idea that knowledge is not an archive but a structure that can be preserved, retrieved and made to operate. Sustenesis adds another requirement, however: the value of memory is not exhausted by query coverage. It also has to preserve identity, boundaries and long-term coherence. Maximising coverage and sustaining a subject are not necessarily the same objective.

Weakness: The formalism depends heavily on the chosen generation operator and definition of coverage. Meaning, conflict and value priority in personal knowledge are still poorly represented. Verdict: Essential.

MindMemOS: A Portable and Self-Evolving Memory Operating Layer for AI Agents

Authors: Kaichao Liang, Yuqi Cui, Hao Kong, Xinyuan Huang, Guohaotian Hou, Qingcan Kang, Liang Chen, Yiyang Yin, Ke Ye, Jiaquan Guo, Da Chen, Lingan Zeng, Yixing Peng, Rong Yao, Shixiong Kai, and Mingxuan Yuan. Published: 12 August 2026. Direct link: https://arxiv.org/abs/2608.12428

Central claim: MindMemOS treats memory not as a fixed database but as a portable operating layer that can reorganise itself. It adapts memory schemas to scenarios, merges redundant records, resolves conflicts and converts execution trajectories into reusable skills. Implicit human corrections are also used to identify inaccurate or misaligned memories.

What is genuinely new: Memory acquires its own lifecycle and evolution mechanism. More importantly, factual memory, pattern discovery and skill formation are connected, allowing “what happened before” to become “how the system should act next time”.

Why it matters for Sustenesis: This is close to the dynamic character of sustained structure. Continuity is not the freezing of an old state; it is persistence through change. But self-evolution immediately creates an identity problem. If schemas, skills and preferences can all be rewritten automatically, on what basis is the system still the same personal AI? Sustenesis can offer a stricter continuity criterion than task performance alone.

Weakness: Evaluation focuses mainly on task accuracy and skill success. Long-term error accumulation, preference drift and the user’s ability to reverse structures formed autonomously receive much less attention. Verdict: Essential.

Agent Safety Should Be a Runtime Contract

Authors: Albus W. Ng, Yi Han, Jusheng Zhang, and Wenhao Wang. Published: 11 August 2026. Direct link: https://arxiv.org/abs/2608.11274

Central claim: Agent safety cannot be treated only as a model property installed during training. An agent that executes code, mutates files, sends messages and modifies databases needs safety contracts enforced by its runtime. The paper gives the contract two faces: preventing dangerous actions in advance and requiring verifiable evidence of successful actions afterwards, such as tests, logs, diffs and grounded citations.

What is genuinely new: The unit of safety shifts from the model to the trajectory with evidence. This resembles permissions, transactions and auditing in ordinary software engineering and is more operational than asking the model simply to be better behaved.

Why it matters for Sustenesis: This is an unusually clear engineering example of constraint. A system remains stable not because dangerous internal possibilities never arise, but because some state transitions are structurally prohibited and others must leave evidence. Constraint is not an external subtraction from capability; it becomes part of the structure that makes sustained trust possible.

Weakness: Runtime contracts can enforce only boundaries that have been formalised. Open-ended ethical conflict and genuinely novel risk can still fall outside the contract. Verdict: Essential.

Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory

Authors: Taeil Kim, Kangsan Kim, and Sung Ju Hwang. Published: 7 August 2026. Direct link: https://arxiv.org/abs/2608.07169

Central claim: Smaller models have difficulty producing enough successful trajectories to learn from their own experience. The authors therefore use a larger teacher agent to distil successful trajectories into three memory layers: workflow, subtask and function-call memories. The layers are injected proactively at task start or reactively when tool-use errors occur.

What is genuinely new: Capability transfer does not have to occur through parameter retraining. It can happen through inspectable external memory structures. The paper reports sizeable gains across several tool-use benchmarks, with subtask memory producing the largest contribution.

Why it matters for Sustenesis: This is a useful technical case for the question of whether knowledge has to exist inside a subject. A weaker model plus an external structure can exhibit a stable ability that the model alone lacks. Does that capability belong to the model, the memory, or the system they form together? Sustenesis is naturally pushed toward the operational whole rather than toward parameters as the only location of knowledge.

Weakness: Successful teacher trajectories can carry their own errors and biases, and benchmark skill transfer is not equivalent to durable open-world knowledge. Verdict: Worth reading.

The Organizational Behavior of Agentic AI: Collective Intelligence in Human-Agent Workflows

Author: Canhui Liu. Published: 29 June 2026. Direct link: https://arxiv.org/abs/2606.30986

Central claim: Multi-agent systems increasingly resemble organisations in functional form: planners, solvers, reviewers, memory managers and orchestrators divide labour and coordinate outcomes. Yet they are not held together by the human organisational foundations of identity, trust, motivation or socialisation. They are sustained by context architecture — prompts, shared state, schemas, permissions, validators and tools. The paper describes the resulting coordination burden as contextual transaction cost.

What is genuinely new: Rather than casually comparing an agent team to a company, the paper specifies where the analogy works and where it breaks. Simulations and trace analysis suggest that copying human organisational forms can add lossy hand-offs and verification burden, while shared-state and adaptive designs can work better.

Why it matters for Sustenesis: Organisation can here be analysed as a structure that continuously coordinates differences and constraints without necessarily presupposing human psychological properties. That does not make an AI organisation ethically equivalent to a human one. Functional organisation, subjecthood and responsibility remain distinct.

Weakness: Context architecture is given a very large explanatory role, while power, interest and institutional history in real organisations are less tractable and receive less attention. Verdict: Worth reading.

The Shift to Agentic AI: Evidence from Codex

Authors: Drew Johnston, David Holtz, Alex Martin Richmond, Christopher Ong, Prasanna Tambe, and Aaron Chatterji. Published: 25 June 2026. Direct link: https://arxiv.org/abs/2606.26959

Central claim: Using large-scale privacy-protecting usage data from Codex, the paper studies the move from conversational AI toward agents that take actions. The authors report more than fivefold growth in active users during the first half of 2026, with use expanding beyond the original software-development population and particularly strong penetration in organisational settings.

What is genuinely new: Claims that AI is moving from answering to acting are often narrative or forecast. This paper provides behavioural evidence from real usage at scale, making the change in work practice observable rather than merely predicted.

Why it matters for Sustenesis: Once AI becomes an action node rather than only a language interface, the boundary of the human–AI system changes materially. Knowledge, judgement and execution are distributed across model, tools, files, permissions and people rather than sequentially contained inside one subject. A theory of the personal external brain therefore has to include continuity of action as well as continuity of memory and text.

Weakness: The dataset comes from one product ecosystem. Rapid adoption does not prove equal gains in productivity, quality or durable capability, and organisational uptake cannot be directly extrapolated to the whole labour market. Verdict: Worth reading.

Synthesis

The six papers form a coherent pattern. Agent research is moving from “how intelligent is the model?” toward “how does the system continue?”. Memory is formalised as a capacity-limited state, built as a self-evolving operating layer and distilled into transferable external skill. Safety moves toward runtime contracts and evidence chains. Multi-agent coordination becomes a form of organisation with its own structural logic, while real usage data show agents entering work rather than remaining in conversation.

For Sustenesis, this is unusually concrete. Difference, Constraint and Sustained Coherence appear in engineering form as heterogeneous states and roles, permissions and runtime contracts, and mechanisms that preserve memory and action across time. Yet a crucial distinction must remain stable: engineering coherence shows that a system can continue to work. It does not by itself generate subjecthood, responsibility or normative legitimacy.

Two questions deserve further development. First, when a personal AI can rewrite its own memory schemas and skills, which changes still belong to the same external brain and which constitute identity drift? Second, when work is jointly produced by a person, model, memory layer, tools and runtime contracts, can responsibility itself be analysed as a structural property, or must normative responsibility ultimately return to a human subject capable of bearing consequences?