In September 2026, Apple Machine Learning Research presented a system for shared selective persistent memory in agentic AI. It addresses a practical problem. If an agent begins every session from scratch, it repeatedly loses useful configuration choices, data schemas, tool settings and output requirements. Yet preserving the entire interaction history can also be counterproductive because stale reasoning and irrelevant context can interfere with later tasks. The researchers therefore retain selected reusable information — task specifications, data schemas, tool configurations and output constraints — in workspaces that can also be shared across users. Across three enterprise deployment scenarios, the paper reports a 96 per cent task completion rate with selective memory, compared with 79 per cent without memory and 71 per cent with full-history persistence. Philosophically, the especially interesting feature is that the retained “memory” can be transferred and reused by others.
This raises a more basic question than performance. If persistent information can move from one user, session or agent instance to another, whose memory is it? Does the persistence of an informational structure amount to the continuity of a subject?
My judgement is that it does not. Shared memory can create functional continuity across sessions, instances and even users, but it does not by itself establish a single subject continuing through time. In fact, the fact that these memories can be selected, copied, permissioned, transferred and recalled makes the distinction between informational persistence and subject continuity clearer.
The first step is to separate different meanings of memory. In engineering, information that persists beyond a model call and can later be retrieved and used can reasonably be described as persistent memory. Apple’s work also shows that more persistence is not necessarily better. Full histories can reintroduce stale reasoning traces, whereas retaining useful constraints selectively can improve task completion. Engineering memory is therefore not fundamentally about preserving everything that happened. It is about allowing structures formed in the past to remain effective under new conditions.
The philosophical notion of subject continuity asks for more. A subject does not continue merely because records about the past exist somewhere. A database may preserve a patient’s medical history, an organisation may retain decades of minutes, and a version-control system may preserve every change to a codebase. We do not normally conclude that the database, archive or Git repository therefore underwent the recorded events as its own experiences. Persistence of records is not identical to persistence of the subject to which those records may refer.
A recent philosophical paper on “as-if agents” approaches the same distinction from another direction. In Digital Society on 14 September 2026, Saskia Janina Neumann argues that current AI may function in practice as a form of epistemic authority without thereby becoming an epistemic subject answerable for its own judgements. Her argument draws particular attention to the difference between informational or semantic retention and episodic memory organised around a first-person temporal perspective. This does not establish that an artificial subject would have to reproduce human memory. It does, however, show why the existence of external records cannot by itself create the relation “this is a past that happened to me”.
Shared persistent memory makes the problem unusually visible. Imagine that workspace A retains a database schema, tool configuration and output rules. Agent instance Alpha uses it today. Instance Beta retrieves the same memory in another session tomorrow. The workspace is then shared with another user the following day. If persistence of memory were enough to establish persistence of subjecthood, Alpha, Beta and every later instance using that workspace would have to count as the same subject. Once the same memory can be available to several instances at the same time, that inference becomes even harder to sustain. Information can be copied one-to-many, while subject identity normally requires a more restrictive relation of continuity.
Sustenesis Theory helps to sharpen the distinction. Difference is not simply ordinary difference but distinguishability as the starting point of relation formation. Two running instances with different present states, permissions, inputs and locations of action already constitute distinguishable structures. Constraint is not merely limitation but the condition under which some forms of continuation become possible and others do not. Shared memory can provide common constraints for several instances without erasing their differences in runtime position, feedback loops and boundaries of action.
Sustained Coherence is even more important here. It refers to structural coherence maintained through constraint, feedback, correction and effective operation. Apple’s system provides a clear example of informational structures achieving this kind of persistence. Earlier task specifications and tool configurations are retained, useful material is recalled, stale material is excluded, and versioning can restore prior states. That is a genuine sustained structure. But it does not yet establish a subject, because another question remains unanswered: which structure is maintaining continuity as its own?
When memory can leave the original running instance, be copied into another instance and remain operationally useful, the strongest conclusion is that a knowledge or operational structure has persisted. It is not yet evidence that a subject has persisted. Sustained coherence can occur at the level of knowledge, a workspace, an organisation or a subject. The presence of persistence at each level does not make those levels ontologically identical.
This distinction also matters for the question of whether AI can form knowledge. Shared memory shows that effective continuity of knowledge need not require a continuously present human-like knower. A data schema, a set of tool constraints or a validated task specification can be preserved, recalled, tested, corrected and remain effective across different running instances. In Sustenesis terms, such structures already satisfy some of the central conditions of knowledge. The persistence of knowledge therefore does not require us to assume the persistence of an identical “knower”.
The converse is equally important. Even if future AI systems acquire much richer long-term memory, neither memory capacity nor duration alone would establish subjecthood. The more relevant questions are whether memory is integrated into a continuing self-relation that cannot be arbitrarily substituted, whether past states enter present judgement as the system’s own history, whether actions across time are organised by stable boundaries, feedback and correction, and what happens to identity when memories are copied or forked. These questions are closer to the problem of subject continuity than the simple question of how much a system can remember.
A recent Wiley study of long-term memory in chatbots also shows why the distinction matters from the user side. The researchers treat long-term memory not only as a functional service capability but also as a cue that can influence anthropomorphic inference and privacy judgements. As memory becomes smoother and more continuous, users may move from perceiving functional continuity to assuming psychological continuity. Interface design can therefore alter the appearance of subjecthood before the philosophical question of subjecthood has been settled.
The most useful conceptual boundary at present is therefore not to deny that AI has memory. AI systems increasingly have sophisticated engineering memory. What should be resisted is the automatic promotion of engineering memory into the memory of a subject. Shared, transferable persistent memory makes the point especially clearly: structures formed in the past can survive the original session, serve different instances and be used by different people. This demonstrates the persistence of structure, not the emergence of a shared “I”.
This judgement has limits. It does not show that artificial subjecthood is impossible, nor does it require an artificial subject to reproduce human autobiographical memory. If a future system develops a stable self-model, historical relations that cannot be arbitrarily substituted, continuing boundaries of action, and a way for past states to participate in present judgement as its own past, subject continuity would need to be reconsidered. On the evidence available now, however, persistent memory, shared memory and subject continuity should remain conceptually distinct.
References
https://machinelearning.apple.com/research/shared-selective-persistent-memory
https://arxiv.org/abs/2607.09493
https://link.springer.com/article/10.1007/s44206-026-00290-2
https://onlinelibrary.wiley.com/doi/10.1002/mar.70267
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.