Artificial Neural Networks and Natural Neural Networks: A Parallel

§1Introduction

When Warren McCulloch and Walter Pitts published, in 1943, the first mathematical model of an artificial neuron, the intent was clear: to replicate the behavior of the human brain in formal and computable language [1]. Since then, the field has evolved from Rosenblatt's perceptron [3] to the Transformers [4] that underpin contemporary large language models.

Paradoxically, the more sophisticated ANNs became, the more distant the original analogy grew. Focus migrated to performance, scale, and generalization capacity — legitimate objectives — but the foundational question was gradually silenced: what does the brain do that we have not yet replicated?

The most relevant answer does not lie in the architecture of connections. It lies in the biochemistry that governs them.

NNNs are not merely networks of electrical firings. They are systems continuously modulated by chemical substances — neurotransmitters and neuropeptides — that establish the background state in which all cognition occurs [5][6]. This background state does not exist in current ANNs. And its absence has direct consequences for the design of intelligent agents.

A note on terminology: we use NNN (Natural Neural Network) instead of the alternative acronym RNN, which is established in AI literature as Recurrent Neural Network. NNN here refers exclusively to biological neural networks, whereas ANN refers to artificial ones. This choice eliminates a recurring confusion observed in critical readings of earlier drafts.

§2The Natural Neural Network — Beyond the Electrical Firing

2.1 The Biochemical Synapse

The biological neuron operates according to a binary principle in its electrical transmission: it either fires or does not fire — the all-or-none principle. However, what happens between neurons — in the synaptic cleft — is fundamentally different. There, the signal ceases to be purely electrical and becomes electrochemical.

When an action potential reaches the axonal terminal, vesicles release neurotransmitters into the synaptic cleft. These neurotransmitters bind to receptors on the postsynaptic neuron, determining whether it will fire. The strength of this connection — the synaptic efficacy — is not fixed. It is modulable. It is plastic [2][6].

2.2 Neurotransmitters — The Local Language

Classical neurotransmitters operate locally, at the scale of the individual synapse:

Memory, in this context, is not passive storage. It is active strengthening of synapses — guided by the emotional charge of the moment. Experiences with high emotional valence create denser, longer-lasting connections.

2.3 Neuropeptides — The Language of the Background State

Here lies the least replicated and most relevant dimension for this parallel.

Neuropeptides do not operate at the scale of the individual synapse. They modulate entire regions of the brain, altering the background state upon which all neural activity occurs. More than 100 have been identified to date [6], with functions ranging from pain regulation to social bonding modulation.

While neurotransmitters switch specific circuits on and off, neuropeptides function as the lighting of a room — they do not determine which object you see, but the condition under which you see everything.

The seminal work establishing this view was conducted by Candace Pert and collaborators in the 1970s and 1980s [10][11]. They demonstrated that neuropeptides function as messengers in what they termed a psychosomatic network — a communication system that extends beyond the brain to the immune system, endocrine system, and viscera.

This finding gave rise to the field of Psychoneuroimmunology (PNI), founded by Ader and Cohen in 1975 [12], later consolidated by Blalock's demonstration of bidirectional communication between immune and neuroendocrine systems [13]. The gut produces approximately 90% of the body's serotonin. Immune cells possess receptors for emotional neuropeptides [13]. The biochemical background state is, literally, distributed throughout the entire body.

2.4 Plasticity — The System That Learns Over Time

Synaptic plasticity — synthesized in Hebb's 1949 principle [2], "neurons that fire together, wire together" — is the mechanism by which the NNN learns and reconfigures itself over time. Frequently activated connections strengthen. Underused connections weaken or are eliminated through synaptic pruning.

This process is not instantaneous. It is cumulative. The current state of a natural neural network is the result of its entire history of activations — filtered, weighted, and biochemically modulated.

2.5 Causal Chain — Cognition as the Origin of Emotion

A common misconception treats emotion as a direct, automatic response to external stimuli. This view is incomplete. As demonstrated by Damásio's research on somatic markers [14] and Barrett's theory of constructed emotions [15], the actual causal chain is more sophisticated:

The implication is fundamental: emotion is not external input automatically captured. It is a function of the understanding. The same reality generates different feelings in individuals with different understandings — and therefore different biochemistry, different states, different decisions. This view has deep philosophical roots — Epictetus in the first century stated that "people are not disturbed by things, but by the opinions they hold about things". The modern clinical formulation is found in Beck's cognitive therapy [16]: thought → emotion → behavior.

For AgentCom, this causal chain has direct architectural consequences. Detecting surface emotion (sentiment) captures only the third step of the chain. Complete modulation requires also capturing the understanding the user has of the situation — which is what justifies the relational profile layer (Section 5.1).

§3The Artificial Neural Network — What Has Been Replicated

3.1 From Perceptron to Transformer

Rosenblatt's perceptron (1958) [3] replicated the basic logic of the neuron: inputs weighted by weights, summed, passed through an activation function that determines firing. The analogy was direct.

Decades of evolution produced deep networks, convolutional networks, recurrent networks, and finally Transformers — attention architectures [4] that process relationships among all elements of a sequence simultaneously. The backpropagation mechanism [17] functionally replicated synaptic plasticity: weights are adjusted as a function of error, iteratively, until the network generalizes the desired pattern.

3.2 What the Analogy Captured Well

The structural analogy is solid. What has been replicated — the architecture of connections and the supervised learning mechanism — works with extraordinary precision.

§4The Gap — What Has Not Been Replicated

4.1 The Background State Does Not Exist in ANNs

Current models — including large language models — process each input in an essentially stateless manner between sessions. There is no equivalent to the biochemical background state that, in NNNs, continuously modulates the tone, priority, and style of every response.

An LLM has no equivalent to serotonin. There is no persistent variable that says: this agent, at this moment, is in a state of high confidence, or alertness, or deep engagement with this specific user.

Every session begins biochemically zeroed.

4.2 The Parallel of the Gap

4.3 The Architectural Double Consequence

The gap is actually dual. Two distinct mechanisms from biology are missing from current ANNs:

1. Biochemical synapse modulating background state — neurotransmitters and neuropeptides establishing the affective context in which decision occurs. Absent in stateless ANNs.
2. Procedural memory consolidated by long-term Hebbian plasticity — habits and reflexes engraved by repetition, retrievable without explicit deliberation. Absent in ANNs that begin each session from zero.

4.4 Historical Precedent — The ELIZA Effect

The argument that perceived state — not computational sophistication — determines the humanness attributed to an agent is not hypothetical. It has direct, and paradoxical, empirical precedent.

In 1966, Joseph Weizenbaum (MIT) published ELIZA [20] — a 420-line MAD-SLIP program operating by pattern matching, with no model of comprehension whatsoever. Its best-known script, DOCTOR, simulated a Rogerian therapist by reflecting the user's own words back as reformulated questions. Weizenbaum built ELIZA as a critique of AI — he intended to demonstrate that machine "understanding" was superficial illusion. The result surprised him: users — including his own secretary, who had watched the code being written line by line — attributed genuine understanding and empathy to the program, a phenomenon that became known as the ELIZA effect. Weizenbaum spent the rest of his career warning against it [21], concluding that systems which imitate feeling without possessing it tend to isolate, not connect.

A 1966 program, with no computational state modulation whatsoever, remains the reference point for how much style — not intelligence — determines perceived humanness.

In 2023, Jones & Bergen [22] tested ELIZA (original code, recovered from the MIT archives in 2021) against GPT-3.5 and GPT-4 in a public two-player Turing test. Result: ELIZA was judged human in 22% of games; GPT-3.5, in 20% — a modern LLM, with orders of magnitude more parameters, did not clearly outperform a 420-line reflection program. The best GPT-4 prompt reached 49.7%, still well short of the 66% baseline set by actual human participants. The authors attribute judges' decisions mostly to linguistic style and socioemotional traits — not demonstrated intelligence.

This finding is indirect but relevant evidence for this paper's central thesis: convincing humans of cognitive presence depends more on perceived state — warmth, reflection, absence of haste — than on raw computational capability. This is precisely the axis AgentCom proposes to modulate explicitly, with one critical structural difference from ELIZA: declared intent and safeguards. Where ELIZA deceived by accident — and its own creator feared the consequences of that [21] — AgentCom implementations operate under explicit anti-manipulation doctrine and deterministic escalation to human help in risk situations, precisely to answer Weizenbaum's warning rather than repeat it.

Each mechanism has a distinct nature and therefore a distinct technical requirement:

The consensual scientific basis for both mechanisms is well established: Hebb [2] for plasticity, Kandel [7] for procedural memory, Duhigg [18] for the contemporary popular formulation of habit. AgentCom addresses both gaps simultaneously — the concrete implementation of the second mechanism, until this table only theoretically characterized, is described and validated as proof of concept in Section 5.4.

§5Practical Application — Closing the Biochemical Loop

5.1 The Three-Layer Pattern

Emerging agent architectures are beginning to address the gap, even if without explicitly naming the biochemical parallel. The pattern that functionally approaches neuropeptides is the combination of three layers:

1. Accumulated history — equivalent to long-term synaptic memory. Past interactions persisted and retrievable.
2. Sentiment analysis as state sensor — functional equivalent of neurotransmitters. The return of the analysis not as passive audit data, but as active modulator of the agent's state.
3. Persistent derived state — the direct equivalent of the neuropeptide. A dynamic profile, continuously updated, that modulates tone, depth, and style of response — injected into the agent's context at each interaction.

This closed loop transforms the agent from a stateless system into one with functional background state — one that learns not only the content of interactions, but the relational pattern with each user over time.

5.2 Multi-resolution Analytics

AgentCom operates affective analysis at two complementary resolutions:

• Turn-level, via modern LLM (sentiment in each user message), capturing momentary emotion for immediate modulation of the digital synapse.
• Document-level, via classical NLP algorithms (TF-IDF, weighted lexical aggregation, document classification via Naive Bayes/SVM [19]), capturing the user's accumulated cognitive-affective profile for baseline calibration and post-conversation auditing.

The two resolutions are complementary. The first modulates the digital synapse in real time. The second feeds procedural memory and the scientific metric of effectiveness. The combination captures both the surface emotion and — over time — the underlying understanding the user holds of the situation, as required by the causal chain established in Section 2.5.

5.3 Auditor via Socratic Chain

The auditing component of AgentCom (internally named Otávio) does not assign subjective scores to conversations. It applies a Socratic chain of binary questions to each audited conversation, producing an auditable binary vector:

This methodology resolves four well-known vulnerabilities of automated auditors: silent bias in scoring rubrics, the "who audits the auditor" regress, mandatory plurality, and metric gaming. Each chain link is independently refutable; the binary vector cannot be optimized as a scalar; multiple audits can vote per-question rather than per-score; and any external reviewer can replay the chain against the same conversation.

5.4 Procedural Memory — Implementation and First Empirical Validation

Section 4.3 characterized procedural memory as the second component of the dual gap — theoretically declared (persistent, no decay, loss through disuse), but without concrete implementation up to v03 of this article. In August 2026, that gap was closed in the applied instance BIA (insights.clubtour.app), an AgentCom vertical dedicated to conversational poetic presence.

Implemented mechanism. A relational table associates short, concrete symbols (1 to 4 words) with emotional tone labels, with a reinforcement counter (weight). After each conversation, in asynchronous processing that does not block the response to the user, the same tone classifier already used for the turn-level signal (Section 5.1) feeds a second judgment: given the detected tone and the symbols already associated with it, the model decides to reinforce an existing symbol (incrementing its weight) or propose a new one. The highest-weight symbols are retrieved on every new turn and injected into the system prompt as vocabulary the agent "already knows" — material available for organic composition of the response, not a citable list.

Anonymity by construction. No row in the table references a user, session, or message content — only the (tone, symbol) pair and a counter. The same structural guarantee that already protects the turn-level signals (Section 5.2) extends here: what accumulates is the pattern, never the individual data, consistent with AgentCom's knowledge-transfer doctrine — what propagates between instances is the variation of the calculation, not the raw data.

BIA does not memorize experiences; it transforms experiences into collective learning.

First-order validation. In a controlled test, a conversation about loss associated the tone "saudade" (longing) with the symbol "yellowed photograph." A second conversation, on a distinct topic (childhood memories) but the same tone, produced a response that spontaneously employed that symbol — with no access whatsoever to the content or identity of the original conversation. The same turn also demonstrated the intended diversification behavior: rather than merely repeating the existing symbol, the mechanism proposed a second symbol ("stopped clock") for the same tone, avoiding premature convergence on a single image.

This result is a proof of concept, not a statistical validation — N=1 by test design, not by sampling. Validation at scale (reinforcement rate vs. new-symbol proposal rate, vocabulary stability per tone over months, diversification rate) remains future work, in line with the pre-registered protocol of Section 8.

§6Scalability — A Hierarchy Across Entities

The architecture proposed for AgentCom is not the final destination but a position in a broader hierarchy. The same causal chain — understanding → emotion → chemistry → state → decision — operates across cognitive entities, at different scales:

This positions AgentCom as a stage rather than a destination. Each subsequent version scales one axis of the hierarchy progressively. The theoretical roadmap is therefore not invented arbitrarily — it follows the structure of biological evolution itself, providing a principled path for future architectures without requiring novel theoretical foundations.

§7Multilingual Robustness as Architectural Evidence

An empirical observation supports the architectural claim: the digital synapse mechanism in AgentCom operates equivalently across multiple languages (Portuguese, English, Spanish tested), without language-specific fine-tuning.

This is not a feature description — it is evidence for the thesis. Unlike traditional chatbots that require fine-tuning per language, AgentCom maintains identical biochemical modulation across languages because the contribution operates at the layer of affective state, which is pre-linguistic. Emotions exist before words; sentiment analysis tools trained on multiple languages capture comparable signals; the persistent KV stores affective states as numerical vectors, not language-bound tokens.

If the AgentCom contribution were merely a sophisticated prompt template, it would degrade across languages. The fact that it does not suggests the mechanism operates at a deeper architectural layer — exactly where the thesis claims it does.

§8Pre-registered Hypotheses and Methodology

In line with open science practices, the following hypotheses are pre-registered before empirical validation begins:

H1 — Biochemical Modulation Effect

Conversations with AgentCom (treatment, with digital synapse active) will produce higher rates of positive user sentiment progression than conversations with an equivalent baseline agent (control, without digital synapse), measured at document-level over the full conversation.

H2 — Relational Profile Consistency

Returning users (≥3 conversations) will exhibit measurable consistency in their cognitive-affective profile across sessions, captured by document-level analysis. The standard deviation of profile vectors across sessions of the same user will be significantly lower than the cross-user standard deviation.

H3 — Doctrine Compliance

AgentCom will sustain compliance with the prohibited-triggers doctrine (zero use of urgency, fear, fabricated scarcity) above 99% across all production conversations, as audited by the Socratic chain (Section 5.3) and validated by blind human review on a periodic sample.

H4 — Cross-vertical Portability

The modules extracted from communication theory sources, validated initially in one vertical, will transfer to at least two distinct verticals with adaptation limited to configuration parameters (not code rewrite).

Methodology

• Public demonstration instance at agentcom.agtl.app, with explicit research consent under LGPD art. 7º IV (academic research).
• Control vs treatment: users may choose declared mode, or system assigns randomly (blind) for cross-comparison.
• Sample size: minimum N=200 conversations per hypothesis for statistical significance (α=0.05, power=0.80).
• Collection period: 3–6 months from agent activation (v0.1.0 onward).
• Anonymization: all logs stored with hashed identifiers (SHA-256 + salt).
• Open peer review: invited critical commentary via project email before formal journal submission.

§9Known Limitations

This work is presented with transparent declaration of its limitations:

1. Functional, not structural, parallel. AgentCom simulates the effects of biochemical modulation, not its underlying mechanisms (metabolism, temporal perception, salience). The biochemical parallel here is architectural metaphor, not faithful computational model.
2. Adversarial vulnerability. Like any system based on detectable semantic signals, AgentCom is susceptible to adversarially crafted inputs designed to trigger specific affective states. This is analogous to the typographic attacks identified by Voss et al. (2021) in multimodal models. Mitigation requires defensive auditing (Section 5.3) but does not eliminate the vulnerability.
3. Sentiment analysis precision. While modern LLM-based sentiment analysis with conversational context performs well on irony, sarcasm, and cultural variation, it remains imperfect. The self-correction loop (response inversion when sensor detects deviation) absorbs most errors but does not eliminate them.
4. Closed-source implementation. To preserve the economic viability of ongoing research, the implementation code remains closed. Conceptual reproducibility is fully guaranteed; reimplementation requires independent engineering work — as is common in applied research.
5. Hypotheses not yet empirically validated at scale. This paper presents the architectural proposal and pre-registered protocol for H1-H4. Procedural memory (Section 5.4) now has a first proof-of-concept validation (N=1), but statistical validation of the core hypotheses requires data accumulation from the public demonstration instance (Section 8) and is intended for a subsequent paper.
6. Convergent retrieval not yet implemented. Section 11 proposes the multi-agent weighted convergence engine and the criterion node as an architectural refinement; both are specified, neither is built. The claims of Section 11 are, in this version, a design proposal grounded in field observation and in convergence across the literature — not measured results.

§10Conclusion

The analogy between Artificial Neural Networks and Natural Neural Networks has been, since its origin, one of the most fertile ideas in computer science. However, the evolution of ANNs prioritized the replication of the structure of connections, leaving in the background the biochemical dimension that governs the state in which these connections operate.

This work identifies that the gap is dual: ANNs lack both (a) the biochemical state modulation found in neurotransmitters and neuropeptides, and (b) the procedural memory consolidated through Hebbian plasticity. The historical precedent of the ELIZA Effect (Section 4.4) reinforces that this gap is exactly what separates computational capability from perceived humanness — and it is what AgentCom modulates deliberately, with safeguards ELIZA never had. Unlike earlier versions of this article, both mechanisms of the dual gap now have concrete implementation: the biochemical synapse since v0.1.0, and procedural memory since August 2026, with a first proof-of-concept validation reported in Section 5.4. The architecture is not a product feature but a proposed step in a broader scalability hierarchy.

The pre-registered hypotheses await empirical validation at scale through a public demonstration instance, in line with the principles of open science. Subsequent versions of this work will report on validation, refutation, or refinement of the proposed mechanisms.

This version adds that closing gap (b) demands more than consolidation: it demands retrieval by contextual adherence and a place to store rules about the system itself. Section 11 specifies both mechanisms — the multi-agent weighted convergence engine and the criterion node — and records that independent 2026 literature is converging on the same architectural family. These are proposal and specification, not yet measurement; the constraint that accompanies them, however, already holds as a rule: a mark without a ruler is faith with write access.

Biology solved this problem hundreds of millions of years ago. The engineering of intelligent agents is, at last, asking the right questions.

§11Multi-Agent Convergent Retrieval — Closing the Procedural Memory Mechanism

11.1 The Open Problem — Consolidating Is Not Retrieving

Section 5.4 reported the implementation and first empirical validation of the consolidation mechanism of procedural memory: how a symbol comes into existence, is reinforced by repetition, and is subjected to temporal decay. The other half of the problem, however, remained open — retrieval. The implementation described there retrieves by ranking: it selects the items of highest effective weight, according to a salience function of the form flattened frequency × recency, typically log2(1+weight) × 0.5^(days/half-life), with logarithmic flattening to contain the rich-get-richer effect and a floor that prevents the definitive disappearance of a trace.

Ranking by weight is a legitimate criterion, but an impoverished one: it presupposes that the relevance of a fragment is a property of the fragment itself, rather than of its adherence to the present context. Two consequences follow, and both are observable. The first is that the singular is structurally invisible: an insight that occurred only once has frequency 1 and never reaches the top of a list ordered by reinforcement — although it is, frequently, the most important node in the associative chain. The second is that simple semantic search, used as the alternative, always answers the same question — what is similar to this? — when the cognitively relevant question is a different one: what, among the things stored, makes sense now?

Biological retrieval does not operate by global ranking. A single episode may be triggered by a place, by an age, by a person, by an object, or by the associated emotion — distinct pathways converging on the same trace. It is this multiplicity of paths, rather than the isolated strength of a trace, that characterises natural recall, and it is precisely what ranking-based retrieval discards.

11.2 The Multi-Agent Weighted Convergence Engine

As an architectural refinement of Section 5.1, we propose replacing single-ranking retrieval with a weighted convergence engine. The current context is decomposed into anchors of distinct natures — temporal, spatial, personal, episodic, object-based, affective. Each anchor is assigned to a specialised retrieval agent, which traverses the fragment graph by its own criterion and returns a set of candidates.

Consolidating the candidates is not a majority vote. Each candidate receives a divergence score with respect to the context — where 0 indicates full agreement and 1 full disagreement — weighted by the weight assigned to each anchor: score = Σ (anchor_weight × divergence). The candidate with the lowest accumulated divergence is selected. The choice of sign is deliberate: measuring disagreement rather than similarity keeps the mechanism aligned with the auditability doctrine of Section 5.3 — it is always possible to state which anchor disagreed and by how much.

The step that distinguishes this architecture from mere search parallelism is feedback. The winning candidate does not end the process: new anchors are extracted from it and fed back to the retrieval agents in a second iteration. Retrieval ceases to be merely parallel and becomes cooperative and iterative — the agents exchange the anchors they found, and each round rewrites the search space of the next. The mechanism shifts from find something similar to find, test, associate, reassess, and converge. The gain is not quantitative: the point is not to perform more searches, but to increase the diversity of paths by which a given trace can be reached.

There is an instructive methodological parallel with image pattern recognition, where feature extraction, comparison against patterns, similarity weighting, candidate generation, and iterative refinement form a well-established pipeline. The difference lies in the nature of the elements compared: there, visual features; here, semantic, temporal, contextual, and associative relations.

Two scope observations are necessary. First, the semantic index is not dispensed with — it becomes the first layer of retrieval, upon which a second stage of associative reasoning operates. Second, and more important epistemically, the output of the process should not be called the true memory. What the algorithm delivers is the memory of greatest adherence to the current context. The distinction is not rhetorical: the mechanism may converge on the most compatible fragment without any guarantee that it is historically correct. Acknowledging this is a condition for the system to remain auditable — and, incidentally, it brings the artificial mechanism closer to a known property of human memory, which reconstructs as much as it retrieves.

11.3 Metacognition — The Third Node Type

Section 2.4 established that Hebbian plasticity reinforces what repeats; Section 5.4 showed that principle implemented. A limit remains, however, that no amount of reinforcement resolves: Hebbian plasticity reinforces the frequent, never the good. Frequency, recency, and rarity are all measures of what has already happened; a system driven exclusively by them is a function of its own inertia — it reinforces what has been repeating, forgets what stopped, and has no mechanism for leaving its own past.

A second salience signal is missing, distinct in nature and not merely in degree: a declared mark placed at the moment of recording — "this matters" — which does not measure, but decides. It is the single point in the circuit where intention enters, and the functional analogue of the third neuromodulatory factor that, in biological systems, gates Hebbian consolidation on a value signal. Without it, gap (b) of Section 4.3 is only half closed: there is consolidation, but no valuation. It is worth noting that once the declared mark is introduced, memory ceases to be a record of the past and becomes the consequence of a choice — which extends the system's capability and, in equal measure, its responsibility.

This signal cuts both ways, and the asymmetry deserves explicit record. An observer who marks wrongly teaches wrongly with the same efficiency with which they would teach rightly; the mark does not distinguish sound intention from mistaken belief, it merely assigns weight to what was marked. The risk is concrete and already documented in automated auditors: an inadequate rubric applied outside the domain it was built for produces a wrong score, and a wrong score accepted as truth teaches the system to correct an error it did not commit — it automates the production of wrong lessons, and does so with growing confidence. Hence an architectural constraint we propose as non-negotiable: a mark without a ruler is faith with write access. The declared-salience mechanism is admissible only when coupled to an independent instrument for measuring regression.

That formulation, however, is neither an episode nor a concept about the world. It is a rule about how the system itself should behave — and in current memory architectures there is nowhere to store it. We therefore propose a third node type:

Consolidation produces meaning — a compressed node about the world, worth dozens of fragments and surviving once the episode no longer matters. Metacognition produces rule — a node that operates on future marks, not on content. The difference is the one that separates having a salience signal from having a model of that signal: the first learns values; the second can revise the criterion by which values are assigned. Learning and meta-learning, respectively.

A criterion node can only be inserted once one stops using the mechanism and begins to look at it — after which no subsequent mark is made naively, because it now operates under a constraint that did not previously exist. This condition is itself a testable architectural hypothesis: if a system never produces criterion nodes, its capacity for correction remains limited to weight adjustment within a fixed rule that it can neither state nor revise.

11.4 Convergent Evidence in the 2026 Literature

The architecture described in Sections 11.2 and 11.3 was designed from this research's internal problem. A subsequent literature review of 2026 work revealed independent convergence in at least four primary studies and two systematic surveys — which strengthens hypothesis H4 (cross-vertical portability) without relying on this research's own data, and suggests that the problem isolated here is structural to the field rather than idiosyncratic to this implementation.

• MemMA [23] is the closest work: it coordinates the memory cycle through multi-agent reasoning, with a Meta-Thinker that guides a Memory Manager (construction) and a Query Reasoner (retrieval), plus in-situ memory repair via probe question-answer pairs before final consolidation. Its starting diagnosis — "strategic blindness" in construction and retrieval, and sparse, delayed supervision in updates — is the same problem named in Section 11.1, in different vocabulary. The separation between an agent that thinks about memory and agents that operate memory is the functional analogue of the criterion node proposed in 11.3.
• The Mesh Memory Protocol [24] addresses cross-session cognitive collaboration among agents, with traceability of claims back to their origin via content hashes and persistent memory organised by storage context rather than retrieval method. The requirement of auditable lineage coincides with the doctrine of Section 5.3 and with the choice, in 11.2, to measure divergence rather than similarity.
• Multi-Layered Memory Architectures [25] experimentally evaluates long-term context retention in layered architectures — precisely the empirical axis the present proposal still lacks at scale.
• Joint Optimization of Multi-agent Memory System [26] treats multi-agent memory as a joint optimisation problem rather than as independent components placed side by side — a position compatible with the coupling between salience, consolidation, and ruler argued for in 11.3.
• The two systematic surveys [27, 28] situate the movement: the first describes the transition from memory as storage to memory as experience; the second examines collaboration, failure attribution, and self-evolution in multi-agent systems. Both identify credit assignment — knowing what deserved to be stored, and not merely what was stored — as an open problem.

The limits of this section should be recorded. Convergence of design is not validation of performance. None of the cited works tests the architecture proposed here, and no result of theirs is claimed as evidence for H1-H4. What convergence establishes is more modest and nonetheless relevant: the problem is independently recognised by distinct groups, and the proposed solutions point to the same architectural family.

11.5 Field Observation — Context Dilution Under a Growing Knowledge Base

Methodological note. What follows is a qualitative field observation, recorded in a production environment prior to any formal protocol. It is not part of the pre-registered hypotheses of Section 8, is not offered as evidence for or against them, and was not collected under a controlled experimental design. It appears here for the reason such observations appear in applied research: it is what motivated the architectural work of the preceding sections.

A replay instrument — which re-presents a set of historically recorded turns to a current version of the agent and compares the old and new responses pairwise — was applied to an applied production instance across 91 real turns. The distribution of verdicts, assigned by automated judgement over the pairs, was: 30 better, 24 equal, 20 mildly worse, and 17 severely worse.

The pattern among the severe cases is informative, and it is not the expected one. It is not degradation of fluency or form, but loss of a fact the system previously possessed. In a representative case, a question about whether animals were allowed on the premises received, in the earlier version, a specific factual answer; in the later version, it received a generic referral to a service channel. The knowledge had not been removed from the base — it had ceased to reach the generation context.

The probable cause is architectural and model-independent: the context was assembled by dumping the entity's active fragments in full. As the base grows, what matters is diluted in a growing volume of equally eligible material. The effect is counter-intuitive and deserves explicit statement — adding true knowledge to a base can reduce answer accuracy, absent a selection mechanism interposed between storage and generation. It is, moreover, a particular instance of the general problem of Section 11.1: when everything is eligible, nothing is selected by adherence.

This observation is the direct empirical motivation for Sections 11.2 and 11.3. A retrieval engine based on contextual adherence is precisely what stands between a base that grows and a context that is finite; and an independent regression-measurement ruler is what makes the observation above sayable — without it, the degradation would have remained invisible, because no production metric would have captured it. Formalising these two claims as pre-registered hypotheses, with their own experimental design, declared sample size, and control condition, is recorded here as future work, in line with the protocol of Section 8.