An annotated reading list and a transfer analysis toward "a single model simulating multiple models."
We trace two lines of evidence — inter-identity memory partitioning in Dissociative Identity Disorder (DID) and the distributed nervous system of the octopus — and extract only the engineering constraints that survive scrutiny.
Search & selection criteria
Search date: 2026-09-27. No lower bound on publication year: the classic works (1992–2013) carry the core mechanisms, while recent works (2025–2026) carry engineering realizations.
Inclusion: only works that answer one of three questions — mode of thinking / evolution & formation / runtime iteration & plasticity. Popular-science claims and items without a locatable primary source are downgraded to secondary material and labelled as such.
Corpus: 28 entries (12 DID / dissociation; 16 octopus / distributed cognition), of which 24 are primary academic sources and 4 are secondary.
The ideas transfer, but what transfers is not the imagery of "multiple personalities" or "the octopus." What transfers is one specific engineering question that each line answers.
The octopus line answers: how to distribute decision authority, and what each part should transmit, when central control is computationally infeasible. It is a set of architectural constraints on bandwidth, timescales and autonomy boundaries — isomorphic to the context-pollution and token-bloat problems of multi-role LLM systems.
The DID / dissociation line answers: along what dimensions memory should be partitioned, and what should be allowed across partitions. It offers not "isolate or not," but a visibility matrix far finer than "all-isolate / all-share," plus a key correction — complete isolation does not empirically exist; cross-identity knowledge transfer is the norm.
Four directly transferable principles
Memory visibility should be graded by memory type, not set uniformly per role: procedural knowledge is nearly fully shared; semantic knowledge may be shared but must carry an attribution label; episodic memory is strictly isolated; role-conditioned interpretation is role-exclusive.
Arbitration should operate on two timescales: low-cost, high-confidence decisions go through a local reflex channel (decided within the role and persisted); only conflict and low confidence escalate to the central coordination channel.
Roles upload only compressed signals to the center (conclusion + confidence + evidence pointer), never the full reasoning chain — a hard bandwidth constraint, not a stylistic choice.
Roles should split dynamically under load triggers, rather than being statically defined up front; the triggers are decidable (number of constraint conflicts, switching density), not aesthetic preferences.
Three things that should not transfer
Do not reify a "role" into a "personality." The schema-mode model in the DID field and the "limits of convergence" paper in the octopus field independently issue this warning.
Do not introduce free-form horizontal discussion between roles. Octopus arms do connect directly, but that is low-bandwidth local coordination, not a global debate.
Do not use popular-science statements such as "the octopus has nine brains" as a premise. The defensible statement: of roughly 500 million neurons, more than two-thirds lie in the arm ganglia outside the central brain.
Part 2 — The DID / dissociation line (12 entries)
The value of this line is not clinical: it is the only long-studied, naturally occurring "single substrate, multiple roles + memory partitioning" system in humans, and so it has accumulated empirical data on "how much isolation is enough, and what isolation costs."
2.1 Mode of thinking — what is actually isolated between identity states
1. Inter-identity amnesia and memory transfer in dissociative identity disorder
Beker et al., 2024 · Clinical Psychology Review (S0272735824001351)
Question
Whether DID patients' reported inability to retrieve other identities' memories holds on objective tasks.
Findings
Using objective memory tasks rather than self-report: identities are not fully partitioned — episodic, semantic and procedural memory all show cross-identity transfer. Dissociative amnesia may relate more to dysfunctional metamemory beliefs (e.g., "recalling trauma will make me lose control") than to a true storage deficit.
Value
The single most important correction to "complete isolation" designs. Isolation in biological systems is never all-or-nothing but graded by memory type; copy that tripartite scheme rather than building a wall.
2. Interidentity amnesia in dissociative identity disorder
Huntjens et al., 2017 · PubMed 28545341
Question
Whether amnesia between alters is real.
Findings
One identity often claims no awareness of events during another's control, but some material types show cross-identity retrieval; the degree varies across individuals, with three topologies — mutual awareness, one-way amnesia, two-way amnesia.
Value
Empirical support that "isolation strength" is a graded continuum and gives three typical topologies that can serve directly as visibility configurations for a multi-role system.
3. DID and memory dysfunction: the current state of experimental research and its future directions
Dorahy, 2001 · Clinical Psychology Review · DOI 10.1016/s0272-7358(00)00068-4
Question
A systematic review of all empirical evidence on DID memory anomalies.
Findings
Using Putnam's (1994) classification of memory anomalies as a framework, it grades the experimental support for each class and distinguishes self-report from objective-task results.
Value
The field's evidence map — useful for quickly judging which claims about isolation have experimental support and which are clinical impressions.
4. The Context-Dependency Model: A Model of Memory Functioning in DID
Macquarie University thesis · figshare archive
Question
The core DID controversy — do multiple identities truly have independent memory streams?
Findings
Proposes context-dependent retrieval in place of "independent memory streams": memory is not stored in isolation; retrieval fails because retrieval cues do not match the encoding context.
Value
An alternative to the "physical isolation" route, highly relevant to engineering — many apparent "role cross-contamination" problems are actually retrieval-cue mismatch, not un-isolated storage. Ablations must distinguish the two causes.
5. Modeling integration and dissociation in brain and cognitive development
O'Reilly et al., 2006 · chapter paper (full PDF via ResearchGate)
Question
What is the computational relationship between integration and dissociation?
Findings
Models integration and dissociation as a computational trade-off in neural networks — separated representations reduce interference but sacrifice generalization; integrated representations do the reverse. Dissociation is not a fault but one solution of a trade-off.
Value
The most important theoretical fulcrum of the line. "Should roles share memory?" is a parameter with an optimum, not a yes/no question; expose the parameter for scanning rather than fixing it at one extreme.
6. A neural network model of dissociative disorders
1992 · Psychiatric Annals / Healio · DOI 10.3928/0048-5713-19920301-11
Question
How to represent memory dissociation in MPD and PTSD with a neural network.
Findings
Proposes that dissociated memory differs in kind from mere forgetting — the former is partitioned and still recoverable, the latter is trace decay. It further treats the realization of consciousness within the network.
Value
A clear design rule: information refused by a write gate should not be deleted but enter a recoverable, auditable diagnostic layer — isomorphic to a LOW-CONFIDENCE branch that only writes a diagnostic log, not the global chain; the paper supplies the mechanistic argument for why.
7. Amnesia and the declarative/nondeclarative distinction: a recurrent network model of classification, recognition and repetition priming
2001 · Journal of Cognitive Neuroscience 13(5):648–669 · PubMed 11506662
Question
Whether the dissociations in amnesia (some learning preserved, some impaired) require a multi-system explanation.
Findings
A single-system connectionist model reproduces all key dissociations — modeling amnesia as a reduced learning rate simultaneously explains preserved classification, impaired recognition, and the recognition/priming split; a non-selective deficit is the better explanation.
Value
A methodological warning. When "role A does not know role B's affairs," do not immediately credit an isolation mechanism — it may be the mundane cause of insufficient retrieval strength. Any isolation-effect validation must be distinguishable from the trivial baseline of lowered retrieval strength.
8. Temporal depth in a coherent self and in depersonalization: theoretical model preprint
2025 · Frontiers in Psychology · DOI 10.3389/fpsyg.2025.1585315
Question
How to characterize the coherence of the self dynamically.
Findings
Uses an attractor-landscape computational model — self-states are attractors; a coherent self corresponds to deep, stable attractors, while depersonalization corresponds to a change in landscape structure.
Value
A measurable language for role switching: switching cost = barrier height between attractors; identity drift = barriers being flattened. Combined with switching-density data, the relationship between switching frequency and barrier height can become a measurable curve rather than a rule of thumb.
9. Towards a general model for psychopathology preprint
2019 · arXiv:1909.02199
Question
Can a unified computational framework explain complex PTSD, DID, schizophrenia and others?
Findings
Uses artificial neural networks to characterize different pathologies as parameterized variants within one framework, treating DID as one case.
Value
Demonstrates modeling "multiple role organizations within one parameter space" — rather than writing separate logic per multi-role scheme, let them be different parameter points of one architecture.
10. Revisiting the etiological aspects of DID: a biopsychosocial perspective
Şar, Dorahy & Krüger, 2017 · Psychology Research and Behavior Management · DOI 10.2147/prbm.s113743
Question
The causal structure of DID.
Findings
Positions DID as a chronic post-traumatic disorder; developmental stress in childhood (abuse, emotional neglect, attachment disruption, boundary violation) is the core, typical cause, alongside family, social and cultural factors.
Value
The key point: partitions form within a developmental window, triggered by overload, not by design. For a system: the role table should not be fixed at design time but split under runtime load (see Increment 4).
11. Modeling the adaptive role of environmental factors and epigenetics in DID secondary
Secondary: ebiotrade report 2026-07-23; original in Cognitive Systems Research (to be located)
Question
How to model DID as a high-order adaptive dynamical system using a self-modeling network.
Findings (as relayed)
A temporal-causal network whose connection strengths self-adapt across five adaptive levels; crucially, the static network parameters are themselves represented as "self-model states," enabling multi-level adaptation. Simulations show prefrontal control going to zero after traumatic stress while adaptive parameters remain high — an irreversible change; shallow interventions improve cognitive symptoms but not deep parameters.
Value
"Represent parameters themselves as state, regulable by a higher layer" maps to: a role's configuration (thresholds, visibility, weights) should be a first-class, monitorable, regulable object, not a hard-coded constant. Also an engineering warning: changing only prompts (shallow) will not change the failure modes of memory structure (deep). Numbers and levels not verified against the original.
2.3 Iteration & plasticity — how to re-integrate after partitioning
12. New evidence-based advances in the treatment of DID (review) secondary
Secondary: ebiotrade report 2025-09-26
Question
The theoretical models and treatment paradigm shift in DID.
Findings (as relayed)
Three main models — the trauma model (incl. structural dissociation), the sociocognitive model (warning of false-memory risk), and the Schema Mode Model, which understands identity states as moment-to-moment emotion–cognition–behavior patterns and, based on cross-identity knowledge transfer, explicitly avoids reifying identity into separate individuals. Neurobiology has not reliably identified DID-specific biomarkers; reversible dissociative amnesia is incompatible with structural brain abnormality, so "dysfunctional metamemory beliefs" may be more persuasive.
Value
The Schema Mode Model is the line's most worth-borrowing conceptual frame — treat "roles" as transient modes rather than persistent entities, matching the implementation where a role is a conditioned call, not a resident object, and preventing unbounded role proliferation.
Additional verifiable leads (not counted in the 12):
EEG-based machine/deep learning for diagnosing dissociative disorders: a review (2025, S2667174325002083) — notes changes in functional-connectivity patterns between host and alter states, a biological echo of "a role = a direction in activation space." Link.
secondary Reinders et al. classify DID vs. healthy MRI with machine learning (32 DID, 43 controls, ~73% accuracy, largest sample at the time); King's Health Partners news; original in The British Journal of Psychiatry. Link.
Part 3 — The octopus / distributed cognition line (16 entries)
This line offers a clear architectural argument: distribution is not a stylistic preference but the necessary solution when the body's degrees of freedom exceed the feasible region of central control. The argument form can be moved directly to decide "when to introduce multiple roles."
3.1 Mode of thinking — hierarchy, local autonomy, bandwidth constraints
13. The Embodied Octopus: Distributed Intelligence and Active Inference in a Flexible Body preprint
Shaw & Berndt (University of Exeter), Jan 2026 · PsyArXiv · DOI 10.31234/osf.io/p7f3k_v1
Question
What computational framework should explain the octopus's distributed nervous system?
Findings
The highest-transfer entry in the list. More than two-thirds of neurons lie not in the central brain but in the highly flexible arms (citing Hochner 2012/2023); this lets each arm perform local perception, prediction and motor control while the central brain coordinates global behavior (citing Hanlon & Messenger 1996). The authors use hierarchical active inference — local generative models run at the body periphery, high-level priors govern goal-directed policy selection (citing Friston 2010, Buckley 2017). The key architectural argument: unlike vertebrates, the octopus must control thousands of degrees of freedom in a soft body, where pure central control is computationally infeasible (citing Flash & Zullo 2023, Kier & Smith 1985), producing a multi-scale division between local and global inference.
Value
Two directly usable items: (1) a decidable trigger — distribute only when degrees of freedom make central control infeasible, i.e., in LLM terms, when a single context window cannot simultaneously hold all constraints; (2) an architectural template — peripheral local generative models plus high-level priors for policy choice: role branches do local generation and verification, the center only handles goals and conflict arbitration, rather than fusing everything token-by-token.
Why are the octopus's neurons distributed this way?
Findings
Argues the distribution is specifically arranged to realize its embodied intelligence — in a jointless soft structure, arms are organized segmentally and neural distribution serves that morphology rather than the reverse.
Value
Establishes the causal direction "morphology determines control architecture." For LLMs: do not first decide "how many roles"; first inspect the task's constraint structure (how many mutually exclusive constraints, how many incompatible knowledge domains) — the number of roles is an output, not an input.
15. The cephalopod specialties: complex nervous system, learning and cognition
Hochner, 2013 · Canadian Journal of Zoology · DOI 10.1139/cjz-2013-0009
Question
The degree of decentralization and its behavioral consequences.
Findings
The nervous system is highly decentralized, especially in the arms; three-fifths of the octopus nervous system lies in the arms; this organization is functional in producing semiautonomous behavior.
Value
"Semiautonomous" is the precise keyword — neither fully independent nor fully controlled. Roles should hold self-determination within an authorized scope, escalating only beyond it — closer to the biological realization than either "each role runs to completion then aggregate" or "the center decides everything."
16. Peripheral control enabled by distributed sensing in an octopus-inspired soft robotic arm for autonomous underwater grasping
2026 · Nature Machine Intelligence · DOI 10.1126/s42256-026-01230-y (IIT team)
Question
Can "sense and decide with the body" be genuinely reproduced in a physical system and control logic?
Findings
The most solid engineering realization. Two strict levels — a bottom local reflex loop: each sensorized sucker has its own controller that triggers negative-pressure attachment on contact with near-zero latency; a top peripheral coordination layer: receives contact-direction information from all suckers, estimates the object's pose relative to the arm, then decides whole-arm bending/twisting/holding. Suckers embed miniature optoelectronic units (~395 mV/N sensitivity over 0–2 N; sensor plates 7.3 mm; shallow-contact direction error < 18°). During grasping, contacted suckers attach within milliseconds; the coordination layer waits for an ~4 s observation window before pose correction. The architecture explicitly avoids streaming all raw signals continuously to the main controller, greatly reducing bandwidth and central-computation dependence, and leaves room to scale — adding suckers or arms does not make the control burden explode.
Value
Three usable points: (1) two timescales — millisecond local reflex vs. second-scale central coordination, distinguished by responsibility not speed: the local does low-risk, high-confidence acts like "grasp," the center does decisions needing global information like "what pose"; (2) compressed uplink — suckers send the already-abstract quantity "contact direction," not raw light intensity: roles should send conclusion + confidence + evidence pointer, not the full thought chain; (3) sublinear scaling — a hard indicator of whether a multi-role architecture is healthy.
18. Octopus-inspired distributed control for soft robotic arms: a GNN-based attention policy with environmental interaction preprint
2026 · arXiv:2603.10198
Question
How to realize a distributed control policy via learning?
Findings
Treats the octopus nervous system as a leading bio-inspired source, using a graph neural network plus attention to model relations among distributed control units and rich peripheral feedback to produce actions.
Value
Gives the representation of relations among distributed roles — graph structure + attention rather than fully connected message broadcast, matching the sparse topology of "roles communicate only with relevant roles."
19. Armed with information: chemical self-recognition in the octopus
Carls-Diamante, 2019 · Adaptive Behavior · DOI 10.1177/1059712319862253
Question
How do parts of a distributed system distinguish "self" from "other"?
Findings
Reviews parts of octopus neurobiology and behavior tied directly to chemical self-recognition — the octopus uses chemical signals to keep its own arms from tangling with one another; the periphery can discriminate self/non-self.
Value
The biological prototype for "role-boundary self-recognition." For a distributed system to work, each unit must be able to decide locally "does this information belong to me," rather than referring every matter to the center — implying trustworthiness checks should be pushed down to each role, not centralized in one verifier.
If cognition is distributed, where should consciousness / subjectivity be located?
Findings
Discusses whether the octopus's cognitive capacities and behavioral repertoire may support consciousness and asks whether "experience" is locatable in a system whose neurons are widely distributed in the periphery. Later work summarizes the view as asking not just "what is the octopus like," but "what is the octopus like in what environment" — subjectivity may not be locatable by hint at a single center.
Value
A philosophical caution for multi-role systems: do not presuppose one "true subject" hidden in some central module. Behavioral attribution may be distributed; do not design an architecture where "the lead role ultimately decides" and assume it is correct by default — make the arbitration mechanism itself auditable.
21. Inverted Cognition: Toward Minds that Begin with Output and Derive Goals Retroactively
2025 · Springer chapter · DOI 10.1007/978-3-032-00686-8_32
Question
Must all cognition proceed "goal first, action second"?
Findings
Takes the octopus as an empirical case of decentralized intelligence producing goal-directed behavior; argues that in such systems behavior arises from interactions among semiautonomous intelligent parts and goals can be derived retroactively.
Value
An architectural option opposite to mainstream plan-execute: on some tasks, letting roles produce local behavior in parallel and then having an upper layer retroactively induce a coherent interpretation may outperform setting strategy first and dispatching — a third path beyond token-by-token fusion and plan-then-execute.
22. Cognition beyond representation: Varieties of situated cognition in animals
Cheng, 2018 · Comparative Cognition and Behavior Reviews 13
Question
What are the different strengths of situated-cognition claims?
Findings
Distinguishes mild and radical versions of situated/embodied cognition, noting that embodied control in octopus arms — or cognition extended into a spider web — adds a component to standard cognitive theory; the mild and radical versions carry different philosophical commitments.
Value
A graded vocabulary for "how autonomous is a role," avoiding conflation of weak and strong claims in discussion; using it in design reviews reduces ambiguity.
3.2 Evolution — why the distributed architecture appeared
23. The Octopus: A Model for a Comparative Analysis of the Evolution of Learning and Memory Mechanisms
Hochner et al., 2006 · Integrative and Comparative Biology (309+ citations)
Question
Relative to simpler molluscs, how extreme were the changes in neuron number and organization?
Findings
Systematically compares the extreme changes in cephalopod neuron number and organization versus lower molluscs and proposes the octopus as a model system for comparative study of learning- and memory-mechanism evolution.
Value
The baseline reference for "how the distributed architecture was selected in evolution," providing a comparative-method template — identify the adaptive significance of an architectural difference by comparison with a simpler homologous system. When evaluating multi-role architectures, pair strictly against "same base + single role," not against a different base.
Albertin et al., 2023 · Current Biology (24+ citations)
Question
The adaptive significance of cephalopod intelligence and large brains.
Findings
Reviews theories on the adaptive significance of large brains, including predator avoidance and shell-loss hypotheses, alongside genomic evidence (e.g., distribution of protocadherin gene clusters).
Value
A list of candidate explanations for "why grow such a costly nervous system." The shell-loss hypothesis transfers best: after losing the shell's protection, the octopus must compensate with flexibility and intelligence — i.e., when a protective layer is removed, the system is forced to develop distributed active control; for LLMs, role partitioning is forced only when the context window (a "protective layer") is insufficient.
25. Grow Smart and Die Young: Why Did Cephalopods Evolve Such Large Brains?
Amodio et al., 2019 · Trends in Ecology & Evolution (165+ citations)
Question
Cephalopods strongly challenge mainstream intelligence-evolution hypotheses — they evolved complex brains and high behavioral flexibility yet are solitary and short-lived.
Findings
The social-intelligence hypothesis and others fail here because cephalopods are mostly solitary and relatively short-lived; their unique life history and ecological context must be considered.
Value
A counter-intuitive reminder — the payoff of architectural complexity need not come from "sociality/collaboration." In LLM terms, the value of multiple roles may come mainly from partitioned management of constraint conflicts, not from "having roles discuss things," consistent with the evidence that multi-agent debate fails.
26. How nervous systems evolve in relation to their embodiment: what we can learn from octopuses and other molluscs
2013 · Brain, Behavior and Evolution 82(1):19 · Karger
Question
The relationship between nervous-system evolution and body form.
Findings
Describes the distinctive "centralization" of the octopus nervous system into three parts and gives the distribution pattern of roughly 500 million neurons across them.
Value
A quantitative schema of neuron distribution; one original source for the often-cited "two-thirds in the periphery."
27. Cephalopod brains: an overview of current knowledge to facilitate comparison with vertebrates
Shigeno et al., 2018 · Frontiers in Physiology (196+ citations)
Question
Why are cephalopod and vertebrate nervous systems regarded as a classic case of convergent evolution, and how far does the analogy hold?
Findings
Systematically reviews current knowledge of cephalopod brains to support comparison with vertebrates; notes the tendency toward distribution in cranial ganglia and discusses the possibility that octopus-arm nervous systems support something analogous to a somatosensory schema.
Value
A methodological exemplar for cross-system analogy — first specify the dimensions of comparison, then judge where the analogy holds and fails. The "octopus → LLM multi-role" transfer must pass the same test or it is metaphor, not argument.
28. Embodied cognitive evolution and the limits of convergence
2025 · PMC12198903
Question
Can convergent evolution be used to infer similar cognitive mechanisms?
Findings
The most important counter-reference in the list. Cephalopods independently evolved large complex nervous systems, flexible behavior, problem solving and tool use, yet with a structure radically different from vertebrates. Beyond the surface similarity of large brains, cephalopods are mostly solitary and short-lived, with no correlation between brain size and sociality/life history; the author criticizes reviews that, despite acknowledging bodily and neural differences, still focus discussion on "species-neutral ability," and argues differences are at least as interesting and informative as similarities — we can perceive behavioral/problem-solving similarities without assuming the same underlying processes. Specific differences: the octopus central brain's motor regions lack the mammalian somatosensory schema; most neurons are in the arm ganglia, which carry a unique chemotactile system based on secondarily adapted neurotransmitters; distributed neural control means an amputated arm continues complex exploratory behavior.
Value
This directly constitutes a methodological constraint on the present question. In transferring the octopus architecture to multi-role LLM systems, the legal use is to borrow its architectural constraints (infeasible degrees of freedom → must distribute; limited bandwidth → must compress uplink; locally decidable → decision authority should be pushed down); the illegal use is to borrow its imagery ("so the LLM should have nine brains," "roles should be free like arms"). The former is a testable engineering argument; the latter is metaphor. Use only the former.
Two verifiable-in-principle points, both from weaker evidence levels and labelled:
secondary Octopuses have RNA editing (recoding) — altering a sequence after DNA is transcribed to RNA to rapidly change protein function and tune neural response to the environment; protocadherin genes help regulate neuron interaction. These are often cited as "molecular-level rapid plasticity," but the original papers were not located here; trace back to the octopus genome paper (Albertin et al. 2015, Nature, DOI 10.1038/nature14668).
secondary"A rule learned by an arm must be integrated by the central brain before it applies to other arms" — local learning does not automatically generalize globally; an explicit integration step is required. If it holds, the transfer value is high (within-role experience does not automatically benefit other roles; gated write-back plus global synchronization is needed, as in a knowledge-graph upsert). It appeared only in Chinese knowledge-base sources; the original experimental paper must be verified before citing.
Verifiable related work: The octopus genome and the evolution of cephalopod neural and morphological novelties, 2015, Nature, DOI 10.1038/nature14668 — the genomic basis of neural and morphological novelty. Link.
Part 4 — Mechanism mapping table
Table 1: biological mechanism → AI mechanism → whether existing work covers it → new increment.
Biological mechanism
Source
AI mechanism
Covered?
New increment
Peripheral local generative models + high-level priors for policy choice
#13
Role branches do local generation/verification; center only arbitrates conflict
Partly (token-by-token central fusion is too heavy)
Explicit hierarchy of delegated authority
Pure central control infeasible at high degrees of freedom
#13
Introduce role partitions only when one window cannot hold all mutually exclusive constraints
No (existing work presets role count)
A decidable trigger
Local reflex (ms) vs. peripheral coordination (s observation window)
Current state: role chains are fully isolated in one work; another splits facts/insights into two tiers. DID evidence gives finer granularity by memory type.
Memory type
Definition
Cross-role visibility
Basis
Procedural
How-to: format conventions, tool-use methods, verification flows
Fully shared, no attribution label needed
Beker 2024: procedural transfer is most stable
Semantic
Facts: objective events, entities, relations
Shared, but must carry an attribution label and timestamp
Companion experimental requirement: treat "sharing degree" as a continuous parameter to scan (the trade-off frame predicts an optimum), rather than fixing it at complete isolation.
Increment 2 — Two-timescale arbitration (from the octopus's local reflex + peripheral coordination layer)
Current state: a single threshold decides whether to persist. Recommended two tiers:
Reflex channel (low cost, high confidence): decidable within the role, persisted directly to that role's chain without alerting the center (cf. millisecond sucker attachment). Example criteria: format compliant, evidence verified locally, no mutually exclusive constraint triggered.
Coordination channel (needs global information): escalate on cross-role constraint conflict, confidence below threshold, or a new fact requiring a write to the global graph (cf. the ~4 s observation window — it waits for a window before deciding rather than reacting immediately; accumulate several turns and arbitrate in batches instead of calling the central verifier every turn).
This also reduces cost: most turns never reach the center, and central-call frequency drops markedly.
Increment 3 — Uplink signal compression spec (from the bandwidth constraint of Nature MI 2026)
Current state: multi-role systems commonly feed each role's full reasoning chain back into the main context. Roles should upload only a four-tuple to the center:
The "evidence pointer" references an anchor triple, not the source text; when the source is needed, the center fetches it by pointer. This matches an existing cognitive-anchor mechanism but elevates it from an internal storage structure to a cross-role communication protocol.
Rationale (three corroborations): Nature MI 2026 states that avoiding continuous raw-signal streaming to the main controller is key to reducing bandwidth and central-computation dependence; context pollution arises from interwoven full histories entering attention; multi-agent discussion can erase up to 72% of an issue's key facts — the free flow of full reasoning chains is the common cause of all three.
Increment 4 — Load-triggered dynamic role splitting (from DID formation + the Embodied Octopus feasibility argument)
Current state: one work lets the model recognize personalities dynamically but gives no "when to split" criterion; the others preset the role count. Recommended triggers (split if any one holds):
Mutually exclusive constraints: role constraints appear in one window that cannot be satisfied simultaneously (cf. adversarial conflict injection and DID's overload-triggered partitioning).
Switching density over threshold: ρ = S/(W−1) ≥ 0.7 (the high-frequency regime under which baseline identity agreement collapses).
Incompatible knowledge domains: the same entity points to different meanings in different contexts (the "Apple the fruit vs. the company" case).
When none holds, do not introduce role partitions; use a single role. This is the discipline from the strongest octopus argument: the legitimacy of distribution comes from the infeasibility of central control, not from distribution looking more advanced.
Optional Increment 5 — Local experience is not automatically globalized to verify
If "an arm's learned rule must be integrated by the central brain before applying to other arms" is verified against the original, a clear rule follows: within-role accumulated experience does not benefit other roles by default; it enters the shared layer only through an explicit integration step (gated write-back + global synchronization). It currently rests on secondary sources; the original experimental paper must be traced before implementation.
Part 6 — Counter-warnings & evidence grading
6.1 Three things not to transfer
1. Do not reify roles. Two independent lines warn alike: the Schema Mode Model explicitly avoids reifying identity into separate individuals in favor of transient modes; the "limits of convergence" paper explicitly opposes sliding from "describing behavior as intelligent" to "assuming a similar form of cognitive process as explanation." The engineering consequence is concrete: treating roles as resident objects causes unbounded role proliferation, per-role state to maintain, and cross-role consistency as a burden; treating a role as a conditioned call avoids it.
2. Do not introduce free horizontal discussion between roles. Arms do connect directly (intramuscular nerve cords), but that is low-bandwidth local coordination, not a global debate. The LLM-side counter-evidence is hard: multi-agent discussion can erase up to 72% of key facts with stance homogenization; debate can make accuracy decline over time, and models tend to prefer agreement over challenging flawed reasoning; majority voting itself explains most of the gains usually attributed to debate. Hierarchical compressed uplink + local autonomy, not horizontal dialogue.
3. Do not use popular-science statements as premises. "The octopus has nine brains" is not a paper conclusion. The verifiable statement: of roughly 500 million neurons, more than two-thirds lie outside the central brain (Embodied Octopus citing Hochner); Hochner 2013 gives the more specific figure of three-fifths in the arms. The two figures differ; cite the source when using either.
6.2 Evidence grading
Grade
Meaning
Entries here
A (primary, peer-reviewed)
Usable directly as factual basis
#1–7, #10, #13–27
B (primary, preprint)
Citable, noting it is not peer-reviewed
#8, #9; #13 (PsyArXiv); #18 (arXiv)
C (secondary relaying)
Only a lead; trace the original before implementation
#11, #12; the two points in §3.3; Reinders MRI ~73%
6.3 A methodological reminder
The result of #7 (JoCN 2001) should constrain your own experimental design: a single-system "reduced learning rate" model reproduces all the dissociation patterns that seem to require multi-system isolation. Any conclusion that "memory isolation worked" must therefore be paired with a trivial baseline — without introducing isolation structure, merely lower retrieval strength / recall counts and see whether the same effect is reproduced. If it is, the contribution of the isolation structure does not hold.
References
Primary academic literature (in order of appearance)
Beker, J.C. et al. Inter-identity amnesia and memory transfer in DID. Clinical Psychology Review, 2024. link
Huntjens, R.J.C. et al. Interidentity amnesia in DID. 2017. link
Dorahy, M.J. DID and memory dysfunction. Clinical Psychology Review, 2001. link
The Context-Dependency Model. Macquarie University thesis. link
O'Reilly, R.C. et al. Modeling integration and dissociation. 2006. link
A neural network model of dissociative disorders. Psychiatric Annals, 1992. link
Amnesia and the declarative/nondeclarative distinction. JoCN 13(5), 2001. link
Temporal depth in a coherent self and in depersonalization. Frontiers in Psychology, 2025. link
Towards a general model for psychopathology. arXiv:1909.02199, 2019. link
Şar, V., Dorahy, M.J., Krüger, C. Revisiting the etiological aspects of DID. 2017. link
EEG-based ML/DL for diagnosing dissociative disorders: a review. 2025. link
Shaw, A.D., Berndt, L.C.S. The Embodied Octopus. PsyArXiv, 2026. link
Hochner, B. An embodied view of octopus neurobiology. Current Biology, 2012. link
Hochner, B. The cephalopod specialties. Canadian Journal of Zoology, 2013. link
Peripheral control enabled by distributed sensing in an octopus-inspired soft robotic arm. Nature Machine Intelligence, 2026. link
Embodying soft robots with octopus-inspired hierarchical suction intelligence. Science Robotics, 2025. link
Octopus-inspired distributed control: a GNN attention policy. arXiv:2603.10198, 2026. link
Carls-Diamante, S. Armed with information: chemical self-recognition in the octopus. Adaptive Behavior, 2019. link
Carls-Diamante, S. Where Is It Like to Be an Octopus? 2022. link