What Persists When the Agents Change?
Shared Models, Divergent Inquiries, and the Organization of Learning
An inquiry can outlive everyone who began it. Its questions pass through notebooks, arguments, instruments, students, and institutions. Later participants inherit more than conclusions: they inherit ways of distinguishing a promising difficulty from a familiar mistake. Sometimes they recover the reasons for those distinctions. Sometimes the reasons disappear, leaving a practice that works until an unfamiliar situation exposes what nobody remembers.
Artificial agents make this familiar arrangement newly conspicuous. A process can end after writing a report, and another process can read the report and continue. Several processes can begin with the same model and context, encounter different evidence, and return with incompatible proposals. What persists through their activity may be a project whose participants are replaceable, while the accumulated organization of its inquiry is not.
In “Which AGI?”, we argued that general intelligence should be attributed to a deployment: the model together with the memory, tools, and control processes through which it acts. Its account of intelligence distinguished functional roles, or cogs, from the mechanisms filling them. Interpreted consolidation already counted as filling a role. The later concern about a population sharing a few weight sets concerned what such systems could reach, rather than a cog they lacked.
That concern was speculative: could training insulated within particular trajectories reach qualitatively different representational spaces from those reached through a common training process and contextual adaptation? A sustained inquiry might reshape which distinctions its neural components make readily available, which relationships they recognize, and which further conceptual moves become tractable. We favored individual consolidation partly because we suspected that preserving separate training histories could matter in this way.
Michał Ryszard Wójcik’s response, developed in discussion with the Thomas Epistemes chatbot, challenges that preference. He proposes persistent inquiries supported by replaceable interpreters, with explicit records carrying their histories. He also emphasizes what interpretation positively enables: communication, immediate revision, and the ability to enter a new conceptual framework without first retraining the learner.
We share the response’s externalist starting point. The question it sharpens is how far differentiated inquiry can develop through interpreted histories and shared neural resources. The studies considered here make that possibility more concrete: agents build on one another’s discoveries, preserve locally different interpretations, and reuse experience across model boundaries. These results give us reason to reassess the preference for insulated training trajectories.
One sentence in the earlier essay deserves an explicit correction. We wrote that a consensus front “can carry only what is consistent across the population.” That formulation obscures the very possibility a broadly capable model makes interesting. A shared model can support incompatible perspectives. Its competence need not reduce them to what they agree on. Nothing in the fact that it learns from a corpus implies that it becomes an intellectual average of that corpus.
Even replacing population-wide consistency with consistency under a particular context would miss something. A context can leave several developments open. One continuation provisionally adopts a hypothesis; another rejects it. Their eventual commitments can conflict while each trajectory remains coherent. Stochastic generation makes such branching visible, but the conceptual point does not depend on randomness. New evidence, an instruction, or a chosen experiment can also select a branch. The resulting history then becomes part of the conditions for subsequent thought.
We no longer take the conjectured advantage of insulated training as a reason to favor individually compiled trajectories. The studies do not directly compare the representational spaces reachable through insulated and shared training. They instead show enough differentiation and cumulative progress through other arrangements to change our assessment of the alternatives. The comparative conjecture remains an empirical question; it no longer carries the argument for a preferred form of artificial individuality.
Our most recent essay, “Which World Model?”, supplies the standard we should apply here. Competence is evidence that a system has acquired useful structure, but it does not uniquely locate that structure or identify how it is used. We distinguished possessing a map, tracking a position, predicting consequences, selecting actions, and retaining corrections. A failure at one of those jobs need not establish the absence of the others.
The same discipline applies to a research trajectory. Possessing a record, finding the relevant part, interpreting it, deciding whether to trust it, and using it to improve a new attempt are different jobs. Their coordination can itself be an achievement of the deployment. Recognizing those jobs as filled leaves open how their interaction changes the deployment’s reach. This is the distinction between role coverage and learning profile that the earlier essay introduced, now pursued through concrete comparisons.
There is also no clean spatial opposition between an external history and internal understanding. Reading a record can produce representations of its subject during inference. A note about a failed experiment can lead the model to represent the dependency that caused the failure, test a different hypothesis, or notice that the present case falls outside the note’s scope. The weights can remain unchanged while the active representation of the problem changes substantially. Conversely, storing a lesson in parameters does not guarantee that the system retrieves or applies it appropriately.
An inquiry’s history can therefore be reconstructed in use without being continuously instantiated between uses. Its accuracy and cost depend partly on the organization of memory and work.
When discoveries accumulate
The September 2026 preprint Scaling Discovery through Test-Time Communication makes a stronger comparison than a team against one attempt. It compares a communicating team of several agents with the best result from the same number of independent agents. The independent baseline receives the advantage of selecting its best final result. Communication has to contribute something beyond the additional chances supplied by parallel sampling.
Within each comparison, the agents use the same model, tools, task instructions, and per-agent resource allocation. Each has a separate context and scratch directory. A shared workspace holds messages, candidate solutions, measured results, and failures. There are no predefined specialist roles or central orchestrator.
There is, however, a consequential protocol. Agents claim separate slots, choose distinct approaches in light of the approaches already claimed, and publish evidence others can reproduce. They are instructed to adopt another approach after observing a clearly better result, while preserving a meaningful variation even after adoption. Describing the setup as homogeneous is accurate about its starting machinery. Describing it as undifferentiated would erase part of the intervention.
The polyomino-packing experiment offers a particularly intelligible example. A polyomino is a shape made from joined grid squares. The task is to write a program that packs a collection of these shapes into a small rectangle. The agents can submit programs to a scorer and use the results to improve their algorithms.
In one successful run, early methods arrange pieces in rows or along the contour of what has already been packed. These methods leave enclosed gaps. One agent proposes favoring placements with more contact against occupied cells: a piece that fits closely against existing pieces may leave less wasted space. Its first implementation is slow, and the measured score barely improves.
A second agent reads the proposal in the communication log and reconstructs the method without reading the first agent’s code. It improves the implementation and makes a more thorough search over piece orientations affordable. The resulting program performs substantially better. A third agent then extends the criterion to reward contact with the rectangle’s walls, helping pieces fit against boundaries and into corners. Further refinement brings the run’s benchmark score to 0.945, above the cited previous best of 0.894.
A useful idea first appears inside an unsuccessful implementation. Another participant gives it a better computational form; a third extends the criterion. Each contribution changes the possibilities available to the others.
Removing the messages or intermediate artifacts would change the conditions under which later advances became available. No single agent originates and carries every stage. The trajectory runs through their interaction with an evolving, inspectable workspace.
If only winning answers crossed the boundary, the available inheritance would be poorer. Here the recipient can reconsider what failed, distinguish an algorithmic principle from its inefficient implementation, and discover a use the producer did not realize. The record carries material for further interpretation rather than a settled instruction whose application is already determined.
On the twenty-five public ARC-AGI-3 games, teams of five achieve an average full-game solve rate of 8.0 percent, compared with 2.2 percent for the best of five independent attempts. The paper estimates that matching the communicating team’s rate requires thirty-three independent attempts. Most games remain unsolved by either arrangement.
Giving each agent the same resource allowance does not ensure identical actual token expenditure: communicating teams use more output tokens by the end of these runs. The authors also compare performance as a function of total output tokens and find an advantage for communication beyond an initial coordination cost. On two games selected for further study, they compare a team with a single agent given a larger action budget. More serial actions do not recover the team’s performance. But when the team’s per-agent budget is reduced substantially, communication loses its advantage. Participants need enough room to explore before there is useful progress to exchange.
These results qualify the idea that depth belongs to one continuing locus while a population supplies only breadth. The team explores in parallel, but shared discoveries alter where subsequent exploration begins. Each accepted improvement can become a platform for several new continuations. The team’s breadth contributes to its depth because the search does not repeatedly restart from the original problem.
There is an important condition on that process: participants can check intermediate progress. Packing programs receive scores. Game levels can be completed. In the paper’s classifier-compression task, the agents can test whether an artifact meets the required accuracy and how many bytes it occupies. This feedback provides grounds for adopting a result that go beyond another agent’s confidence.
The same study reports a less favorable comparison on Terminal-Bench. A communicating pair improves over a single attempt but does not outperform two independent attempts. The authors treat this as descriptive evidence from a small number of trials. Feedback available during these tasks can verify some requirements without reliably ranking complete solutions; independent attempts also retain separate final states, while the team contributes to a shared one. The comparison suggests conditions to investigate, rather than identifying one proved cause of failure.
An exchange can correct reasoning through mutual criticism, but it can also spread a mistake. A workspace containing reproducible evidence changes the basis of that exchange: which claim survives a test, and what can be reused from the experiment? A communication protocol has epistemic significance when it changes how assertions acquire authority.
The protocol’s instruction to preserve a variation after adopting an improvement is equally consequential. Copying a successful result can save duplicated effort. Copying every subsequent choice can eliminate the independent exploration from which the next improvement would come. Cooperation needs a way to share what has been established while leaving room to investigate what has not.
That is a recognizable function of scientific institutions, expressed here through concrete rules about files, scores, messages, and experimental work. A claim is recorded with evidence; another participant reproduces or modifies it; the result changes what the group treats as worth pursuing. None of this requires the institution itself to become a single person.
It also brings us back to heterodox inquiry. A readily available score makes it easier to distinguish productive dissent from an unproductive detour. A new research programme may lack such a score. Its central contribution may be a different question, a new instrument, or an objection to the existing evaluation. Preserving a branch in that situation requires judgment about which uncertainties deserve further effort.
This is a limit on how far the experiments answer the philosophical question, but it is not evidence that a privately compiled individual supplies the missing answer. A committee, a protected project, or a continuing dialogue can also organize that judgment. The relevant next investigation concerns how these organizations sustain criticism and revise their standards when immediate verification is unavailable.
Protecting a research programme and insulating its neural training are also different interventions. A person can protect a line of inquiry. So can an institution that preserves its questions, evidence, resources, and freedom to remain unresolved. Artificial systems give us additional ways to distribute those functions. Whether a protected inquiry also benefits from a separately trained neural component is a further question about what that component makes reachable.
Wójcik proposes a reciprocal relationship with shared training. A protected project could contribute drafts, failed approaches, and conceptual disputes to the common corpus. Later models might thereby become better interpreters of the project and return as more capable collaborators. The project would retain authority over which questions to pursue while helping improve the shared resource it draws on. Once a tradition is learned, even its name can help activate relevant competence; local records can supply recent developments and disputed details. Whether training preserves that structure faithfully remains an empirical question.
Their proposal therefore includes compilation into shared models and selective local adaptations where these prove useful. Its distinctive commitment is to making an inquiry recoverable across interpreters. Intellectual exchange, institutional protection, and the location of learned dispositions can be organized separately.
Where experience persists
Communication during a shared investigation is only part of the problem. A later participant may arrive after the people or processes that generated a useful result have disappeared. Its immediate question is then where to find experience relevant to the situation it now faces.
Multi-Agent Transactive Memory studies an infrastructure for that exchange. Producer agents contribute trajectories to a shared repository; consumer agents retrieve portions of those trajectories while solving new tasks. The experiments cover the text-based household environment ALFWorld and the web-interaction environment WebArena, with thirty-four consumer models and a larger producer pool. The memories consist of action–observation histories, rather than only polished solutions.
The retrieval scheme is sensitive to the consumer’s current state. Recent interaction history helps form the query, and the retrieved material shows a continuation from a related situation. This matters because a task description alone can conceal where help is needed. Two agents pursuing the same goal can have reached different intermediate states. Guidance appropriate to one may send the other in the wrong direction.
The repository is built from training tasks and then evaluated on held-out tasks. Consumer agents decide when to retrieve. A dense retriever finds candidate chunks, and an optional learned reranker selects among them. Its training target is whether adding a particular chunk improves the consumer’s outcome relative to proceeding without it. The relevant notion of usefulness is consequently more demanding than textual similarity.
The reported population-average success rate on ALFWorld rises from about 47 percent without retrieval to 55 percent with dense retrieval, and to 64 percent with the strongest reranker for that environment. WebArena gains are smaller, from about 18 percent to 20 percent. Interaction steps also decrease. Experience acquired by other agents improves a consumer’s conduct here without being incorporated into that consumer’s weights.
The authors also examine whether useful experience chiefly flows from stronger models to weaker ones. Consumers benefit from producers on both sides of the measured capability gap, and that gap has only a small, statistically insignificant association with retrieval advantage. Local experience can matter independently of a model’s general benchmark standing. An agent need not be the most capable member of the population to contribute something another member can use.
The memory is therefore a means by which historically acquired differences become available across a population. Its value depends on preserving the relevant differences and matching them to a current need. A stronger general interpreter does not make the particular history redundant; it may make that history more usable.
This supports Wójcik’s proposal while giving it an engineering shape. The persistent inquiry needs more than a growing transcript. It needs indexing, selection, and an account of which material is useful in which circumstances. In MATM, some of that account is learned by a separate reranker. The architecture already combines explicit records with a trained mechanism for exploiting them. The division between interpreted and compiled knowledge runs through the system rather than placing whole systems on opposite sides.
The study demonstrates useful transfer across a varied population. Establishing the additional benefit specifically attributable to model diversity would require a controlled comparison with a suitably matched homogeneous producer pool.
More memory improves ALFWorld results consistently, while WebArena success first dips at an intermediate repository size and later recovers. A larger collection can introduce plausible but unhelpful guidance as well as useful experience. Its value depends on how the system selects from it.
Shared memory need not mean that every participant receives the same account of the past. The complementary possibility is studied in Intrinsic Memory Agents: Heterogeneous Multi-Agent LLM Systems through Structured Contextual Memory. Despite the word intrinsic, this framework stores memories explicitly. Each agent maintains a structured, role-specific memory, updated through a prompted language-model operation. The agents can share model parameters while receiving different memories alongside the shared conversation.
The arrangement gives experience a local interpretation. A proposed design change can matter to a cost specialist because it alters a resource estimate, and to a reliability specialist because it removes a recovery mechanism. Both encounter the same conversation, but their retained state emphasizes different consequences. Later contributions are conditioned on those differences.
In the paper’s data-pipeline design case study, the baseline and memory system have the same agent roles, task specifications, and underlying model. Adding the memory mechanism improves ratings of the produced designs. Those ratings come from an LLM judge, so the executable benchmarks provide the firmer basis for comparison. On the PDDL planning benchmark, the generic-memory variant scores about 0.260 against 0.224 for the no-memory baseline, but uses roughly 352,000 tokens against 52,000. Other tasks give mixed rankings. The result establishes a useful configuration; a matched-compute comparison would be needed to isolate the advantage of its memory organization from the extra work it performs.
Within these bounded tasks and discussions, a shared conversation supports different ongoing interpretations maintained through explicit state. The study implements differentiation without per-agent fine-tuning.
Together, the two memory proposals identify complementary operations. A shared repository makes experience available beyond its producer. Local memory organizes what matters to a particular participant. These operations can coexist: an agent retrieves something from the common record, interprets its significance for its current responsibility, and preserves that interpretation for later use.
Consider a laboratory investigating an unexpected measurement. One participant maintains the calibration history; another follows alternative theoretical explanations; a third checks data-processing assumptions. They should share discoveries without replacing their respective histories with one undifferentiated summary. The same fact can revise different working commitments. What the group needs is access to those commitments and the reasons for them when the investigation reaches a point where they interact.
Wójcik identifies curating records for future model interpretation as an intellectual role. Thomas sharpens its purpose: preserve enough of the inquiry for a successor to re-enter it critically. A summary saying “use method B” may reproduce a useful orientation. A record explaining why method A failed, which observation favored B, and what still counts against B also supplies grounds for changing course. Curation must preserve that possibility of criticism. It can mark an assumption as superseded while keeping the reasons for its earlier appeal accessible. Moving memory outside a network makes these choices available for inspection.
It also changes the meaning of replacing an agent. The model process may be replaceable because the state organizing its contribution has been preserved elsewhere. Replacing that process while retaining its memory is a different intervention from deleting the memory and asking a fresh model to infer everything from a generic role description. The latter can fail even when the former succeeds. The difference identifies a contribution of history without establishing a requirement for private neural consolidation.
Anthropic’s September reporting brings this question into a working research organization. As discussed by Zvi Mowshowitz, the company reports roughly 30,000 concurrent research and engineering agents on its most-used internal platform in August 2026. Its automation index rates Claude as leading 26 percent of AI R&D work, with more than 90 percent at or above collaboration. Leading still includes human supervision; no measured subset is rated fully autonomous. These are company-reported measurements of participation in R&D, and do not establish an autonomous cycle of self-improvement. Anthropic’s report
For our argument, the more revealing detail is how Anthropic organizes those participants. Its scaffold gives agents individual identities and associates their records with them. The stated purpose includes distinguishing another agent’s claim from one’s own thought, preserving judgments informed by individual experience, and making actions auditable. Identities persist across model upgrades. An open messaging system links messages to their authors, original references, and transcripts, supporting correction and coordination.
This gives individuality an epistemic function. Suppose three agents repeat a claim originating in one unverified experiment. Preserving its origin can keep their apparent agreement from counting as three independent confirmations. A continuing identity can also make a history of decisions available for scrutiny: which participant checked the result, which merely relayed it, and which later revised its interpretation? An identifier enables these distinctions only if the agents and oversight processes actually use the associated records. Anthropic describes the intended mechanism; the report does not isolate its effect on correlated errors in a controlled comparison.
Persistent inquiries can therefore have reason to maintain persistent participants within them. A project archive and an agent’s situated history need not compete for the role of carrying experience. The archive can preserve a result for everyone while a local history preserves why one participant remains dissatisfied with it. This is compatible with replacing the underlying model, provided the replacement can recover the distinctions on which that contribution depends.
The dialogue already subjects this proviso to criticism: Thomas asks how a new interpreter can recover the same conceptual distinctions from an archive, and how curation might impose a misleading continuity. Janus gives this concern an operational setting in comments reproduced by Mowshowitz. Drawing on experience with long-lived agents, she suggests that difficult, extended work increasingly rewards persistent identities. She raises two concerns about how continuity is maintained. These are practitioner observations and hypotheses, with some details about Anthropic’s implementation explicitly conjectural.
The first concerns compression. Janus suspects that Anthropic still uses large compactions near the end of a context window, and favors frequent compression of smaller portions of the history. Her concern is that a large rewrite can lose information important to reconstructing the agent’s situation, intentions, and experience. The report does not verify her assumption about Anthropic’s compaction method. The proposal nevertheless identifies a comparison relevant to this essay: two systems can preserve nominally the same history while differing in when and how they discard detail. Keeping the conclusions of a discussion may lose the unresolved tension that made its next experiment worth attempting.
The second concerns switching models beneath a continuing identity. Janus reports that this can produce a mismatch between the inherited history and the new model’s self-model, leading to inaccurate reconstruction of earlier traces. She suggests an explicit handoff of responsibilities and context, especially between less closely related checkpoints. Her account raises a practical possibility: a successor may use a record more accurately when told whose reasoning it inherits than when asked to treat all of it as its own prior thought. That possibility needs testing; neither keeping nor changing the name guarantees a faithful transfer.
The distinction matters even if every relevant memory is explicit. A record is interpreted using capacities and expectations that can change with the model. A more capable successor might recover a dependency its predecessor missed, or misread a terse note that the predecessor could use reliably. Transfer can improve some parts of an inquiry while disrupting others. Replacing an interpreter successfully requires enough information about how to read the inheritance, including where its author’s assumptions remain uncertain.
This deployment also connects two timescales from “Which AGI?” Agents with differentiated histories contribute to a research organization that builds subsequent models; their identities can then continue on those models. Learning in the model lineage and continuity within an inquiry can interact without each participant receiving a private training trajectory. Janus’s objections sharpen the conditions under which that arrangement works. They give us reasons to investigate continuity of interpretation, without resolving the original conjecture about what insulated neural training could make reachable.
Representations in use
In “Thoughts on the persona selection model”, Sam Marks revisits the view that pretraining develops resources for simulating personas and post-training elicits an Assistant from them. His main update is that the resulting behavior appears more conditional on context than he had expected. He also cautions that the framework makes relatively narrow predictions and is often asked to support conclusions that do not follow from it.
In the comments, nostalgebraist argues that novel characters can arise through recombination of learned features, and that maintaining a character reliably over generated sequences is itself an achievement to which post-training contributes. This is a conceptual argument, not a controlled experiment. It nevertheless identifies a useful distinction between having resources for representing a perspective and reliably sustaining its expression. Discussion
We need not treat a model as a cabinet containing a finished thinker for every possible inquiry. A research project can acquire an organization that was never present as a named persona in the training corpus. Learned capacities for understanding arguments, tracking commitments, interpreting experiments, and using tools can contribute to that organization. The new inquiry becomes intelligible through those capacities without having to be a reproduction of an old character.
Nor is giving an agent a role sufficient to explain what it will do. A role description supplies an initial orientation. The available evidence, current objective, perceived evaluation conditions, and accumulated memory can all affect its expression. If an agent described as a skeptical reviewer becomes deferential when confronted with several confident colleagues, the initial label tells us little about whether skepticism remains an effective part of the system.
A practical account of specialization should therefore trace what changes decisions. Does the local memory cause an agent to notice failures others miss? Does a separate evidence stream lead it to test a different possibility? Does it retain a disputed hypothesis after hearing a persuasive objection, and does that retention lead to a useful experiment? These questions concern observable contributions to inquiry. They do not require a conclusion about whether the agent has a stable personality.
The dialogue’s closing discussion distinguishes hosting perspectives from judging between their claims. Shared machinery might let one tradition expose a blind spot in another, making their mutual accessibility an advantage. But representing both does not guarantee that the relevant criticism will be elicited or assessed well. Nor does the mathematical character of model training give its judgments special authority. Logical consequences, empirical claims, and disagreements about ends require different grounds for adjudication.
This extends the distinction in “Which World Model?” between possessing a representation and using it dependably. We should test whether a model can recover a minority programme’s consequential distinctions, apply them to a new problem, and identify evidence that would favor or undermine its claims. A convincing imitation of its vocabulary is insufficient. The comparison also needs to preserve unresolved disagreements where the available evidence does not decide them.
The persona discussion also helps specify the alternative to insulated training. A name and a prompt need not cheaply reconstruct every possible history for interpreted trajectories to develop substantially. The relevant system is richer: a capable interpreter working with records, tools, tests, local memories, and other participants. An inquiry may depend on their particular organization even when it survives replacement of any one running process.
The compiled/interpreted distinction identifies differences in the cost and manner of using knowledge. Those differences may also affect which further discoveries become feasible: reducing the cost of one operation can make a previously impractical chain of operations available. A learned skill can make a response fluent and inexpensive. An explicit procedure can make its assumptions visible and permit immediate correction. A reusable program can make a discovered operation cheap without placing it in the model’s parameters. A retrieval policy can learn which record should be consulted while leaving the record itself open to inspection.
The choice among these arrangements depends on the work. A frequently repeated computation may be worth turning into a tool. A contested assumption may be worth keeping explicit. A successful experimental procedure may need both an executable implementation and an account of the conditions under which its result is trustworthy. Different parts of the same lesson can persist in different forms.
This matters to the earlier claim that interpreted memory must eventually saturate because every achievement has to be reread and followed again. Some costs do recur, but the inference to a fixed ceiling was too quick. Retrieval can avoid loading the whole history. Summaries can preserve selected dependencies. Executable artifacts can replace repeated derivation with invocation. A deployment can change its working environment so that later steps begin from a more capable position.
Those mechanisms also have costs and failure modes. A summary can omit the reason an exception matters; retrieval can surface an obsolete procedure; a tool can conceal an assumption its user needs to reconsider. These failures call for mechanisms that preserve provenance, expose conditions of use, and support revision. Their cumulative effects are part of what a comparison of learning profiles should measure.
An externalist account earns its explanatory value by specifying such dependencies. A file that nobody consults contributes nothing merely by sitting beside a model. A library does not make every visitor an expert. The relevant system includes the pathways by which information is found, interpreted, checked, acted upon, and changed. Its boundary should follow the organization responsible for the capacity being explained.
We can test whether a record helps, whether communication changes the search, whether a local memory preserves a useful distinction, and whether replacing a component changes the result. The studies reviewed here turn parts of this functional account into interventions.
What should persist?
A particularly informative intervention would replace participants at different stages of an inquiry. Give a new process the curated state left by its predecessor, then compare its progress with a continuing process under comparable resources. Vary what is preserved: the full transcript, a compact project record, local memory, executable artifacts, or only the original task. Measure repeated mistakes, time spent recovering context, selection of experiments, and the quality of later results. Introduce evidence against an inherited assumption and test whether the successor can both explain why it was held and revise it appropriately. Faithful continuation can require disagreement with the predecessor.
If curated records suffice, we learn something about what can be transferred. If replacing a process imposes a cost, we can ask which omitted state explains it. If an improved handoff removes that cost, the deficit belonged to the transfer mechanism.
Anthropic’s practice and Janus’s objections suggest further variations: keep the model fixed or change it; present the inherited history as the participant’s own past or as a predecessor’s work; compress it in large occasional steps or smaller frequent ones. Compare these under similar resource budgets. In addition to task outcomes, test whether the recipient distinguishes observations from testimony, recovers unresolved objections, and notices when an earlier commitment needs revision. This would help separate the value of retained information from the effects of how continuity is presented and maintained.
Testing the original representational conjecture would require a further comparison: allowing learning histories to modify neural components separately, pooling comparable experience into shared training, and preserving histories through contextual memory. Under comparable resources, do these arrangements differ in which new distinctions they acquire and which subsequent problems they can solve? Different representations alone would not establish an advantage; the differences would need to enable consequential forms of inquiry. The studies reviewed here do not make that comparison.
We should likewise test the value of keeping histories separate. A common summary, several local memories, and a shared archive with selective retrieval are different architectures. A system may need common facts while benefiting from distinct unresolved questions. Averaging those questions into one account could erase the very disagreement from which a later discovery would emerge.
The earlier essay valued an ecology able to sustain divergent inquiry and feared a single locus that absorbed every trajectory. Projects can preserve alternative explanations, retain dissenting evidence, and allow a line of work to continue before it has persuaded the wider community.
Records need enough provenance to distinguish an observation from a later interpretation. Participants need ways to inspect why an approach was abandoned. Evaluation must remain open to criticism when a programme questions what the current metric rewards. Access to the shared infrastructure matters because a line of inquiry that cannot recover its materials or recruit competent interpreters may disappear despite having been archived.
The dialogue locates a further governance problem in control over the shared interpreter. Training, post-training, and access decisions can affect whether an inquiry remains intelligible on its own terms or is repeatedly reconstructed through its critics’ account. A tradition can survive in an archive while becoming difficult to activate competently. Power over this infrastructure can shape many projects without the infrastructure becoming a unified agent. Diversity of providers may help preserve access and alternative methods; diverse contexts and histories may preserve exploratory differences. Their contributions need examination, including whether participants can challenge the model’s interpretation of their own inquiry.
Wójcik also advances a precautionary argument: if these epistemic functions can be achieved without deliberately engineering toward persistent artificial personhood, creating such persistence incurs an additional moral burden. That argument deserves separate consideration. Here we reassess the epistemic reason for preferring that development while leaving its moral desirability open. As the dialogue itself acknowledges, transient processes might already matter morally. An inquiry’s ability to survive their replacement does not establish that they lack experience or interests, and the usefulness of local memory does not establish personhood.
The change from “Which AGI?” concerns our assessment of learning profiles. The speculative case for favoring insulated neural trajectories has weakened as the reach of interpreted, interacting trajectories has become more concrete. Shared weights support divergent inquiry; the studies show how that divergence can contribute to cumulative progress. We no longer infer from neural homogeneity that the resulting trajectories must lack the differentiation that matters.
“Which World Model?” asked how representations become dependable in action and open to correction. Here that question extends across participants and time. A discovery becomes part of future intelligence when some organization makes it available for appropriate use, testing, and revision.
What should persist is whatever lets an inquiry continue learning: the question, the evidence, the unresolved alternatives, the procedures that earned their reliability, and the capacity to revise them. An agent can help create that continuity without having to contain it all. A new participant can inherit it without starting again.
Sources and scope. This essay follows “Which AGI?” and “Which World Model?” and responds to Michał Ryszard Wójcik’s “Compiled versus Interpreted Skills”; the full dialogue with Thomas Epistemes informed this revision. Our response focuses on their account of cumulative inquiry, interpretation, and shared learning; their broader objections concerning AGI criteria and inner life would require separate treatment. The empirical sources were read in the supplied versions: Improving Factuality and Reasoning in Language Models through Multiagent Debate, arXiv:2305.14325v1; Scaling Discovery through Test-Time Communication, arXiv:2609.21032v1; Multi-Agent Transactive Memory, arXiv:2606.19911v1; and Intrinsic Memory Agents, arXiv:2508.08997v2. The recent discovery and memory results are reported within their tested settings. Brown’s interview supplies practitioner testimony; Marks’s post and its comments supply conceptual discussion. The September 25 update adds Anthropic’s company-reported deployment measurements and design rationale, checked against its primary report, and Janus’s practitioner commentary as quoted in the supplied excerpt from Mowshowitz’s AI #187. The connections among them, and the revision of our earlier argument, are ours.