metadata (normative)

Title:	Cognitive Memoisation: Governing Knowledge Round-Trip to Prevent Knowledge Erosion in LLM Systems
Curator:	Ralph B. Holland
Affiliation:	Arising Technology Systems Pty Ltd
Contact:	ralph.b.holland [at] gmail.com
Version:	1.0.1
Publication Date:	2026-03-07T08:08Z
Updates:
Binding:	final

The metadata table inthis section is CM-defined and constitutes the authoritative provenance record for this artefact.

All fields in that table (including artefact, author, version, date and reason) MUST be treated as normative metadata.

The assisting system MUST NOT infer, normalise, reinterpret, duplicate, or rewrite these fields. If any field is missing, unclear, or later superseded, the change MUST be made explicitly by the human and recorded via version update, not inferred.

As curator and author, I apply the Apache License, Version 2.0, at publication to permit reuse and implementation while preventing enclosure or patent capture. This licensing action does not revise, reinterpret, or supersede any normative content herein.

Authority remains explicitly human; no implementation, system, or platform may assert epistemic authority by virtue of this license.

Cognitive Memoisation: Governing Knowledge Round-Trip to Prevent Knowledge Erosion in LLM Systems

Scope

This is the Cognitive Memoisation (CM) Purpose Paper.

Anchor

Cognitive Memoisation governs the round-trip of knowledge between humans and stateless AI systems in order to prevent knowledge erosion and preserve human knowledge across sessions and over time.

Abstract

Large Language Model (LLM) systems are inherently stateless. Each interaction is processed within a bounded context that is routinely truncated, paraphrased, or reinterpreted. As a consequence, knowledge introduced during a session erodes through semantic drift, constraint weakening, paraphrasing, or context eviction. Over extended interactions this produces the familiar "Groundhog Day" condition, where previously established facts and constraints must be repeatedly reintroduced in order to maintain progress.

Cognitive Memoisation addresses this problem by externalising knowledge into durable, human-governed artefacts that can be serialised, preserved, and reintroduced into inference as required. Rather than relying on fragile conversational continuity, Cognitive Memoisation establishes a governed round-trip between human cognition, durable knowledge artefacts, and LLM inference. In this model, knowledge is explicitly captured, preserved from erosion, and projected back into reasoning contexts in a controlled and repeatable manner, allowing human knowledge to persist across sessions and over time.

The purpose of Cognitive Memoisation is not to replace human reasoning nor to embed authority within AI systems. Its purpose is to ensure that knowledge produced through human-AI collaboration remains stable, corrigible, and progressively accumulative. By governing the round-trip of knowledge between humans and stateless AI systems, Cognitive Memoisation allows work to compound across sessions without loss of meaning, constraint, or intent.

1. Knowledge Erosion in Stateless LLM Systems

Large Language Model (LLM) systems operate within bounded inference contexts. Each interaction is processed using only the information present in the current prompt and the retained context window. When that context is truncated, rewritten, or replaced, previously established knowledge is no longer available to the model. As interaction continues, information that was previously established is progressively lost from the active reasoning space.

This condition produces what users often describe using imprecise terms such as hallucination, drift, or conversational breakdown. These labels describe observable symptoms but they do not explain the underlying structural pressures acting on the system. In practice the behaviour arises because stateless inference systems cannot reliably preserve knowledge across extended interaction sequences.

During prolonged interaction several forms of knowledge erosion occur.

First, semantic drift occurs when previously established meanings are gradually reinterpreted or paraphrased until the original constraint or definition is weakened.

Second, constraint erosion occurs when explicit rules or invariants introduced earlier in the interaction are no longer applied because they are no longer present in the active inference context.

Third, context eviction occurs when earlier material is removed from the context window due to token limits or prompt restructuring.

Fourth, Groundhog Day repetition occurs when knowledge that was previously established must be reintroduced because it has been lost from the model's reasoning context.

These behaviours are not random faults. They are structural consequences of stateless inference operating across a sequence of bounded contexts. When knowledge must persist across many interactions, the system gradually loses the information required to maintain consistent reasoning.

Describing these behaviours using general terms such as hallucination is therefore insufficient. Such terms group together multiple distinct pressures and obscure the mechanisms responsible for knowledge loss. In order to analyse these conditions more precisely a structured diagnostic model is required.

The Governance Axes multi-dimensional lens provides such a model. Rather than treating failure conditions as isolated anomalies, the lens identifies the governance dimensions under pressure during interaction. Each axis represents an independent governance property such as authority, constraint enforcement, attention, temporal coherence, or normative fixity. By observing which axes are under pressure during interaction it becomes possible to explain why knowledge erosion occurs and how it propagates through a reasoning sequence.

The Governance Lens therefore converts vague descriptions of model behaviour into a structured analysis of governance pressures. This allows erosion to be understood not as a single phenomenon but as the combined effect of multiple pressures acting simultaneously on knowledge, constraints, attention, and temporal continuity within stateless inference systems.

2. Governance Pressure in LLM Interaction

Knowledge erosion in LLM systems is often described using broad or informal terminology such as hallucination, conversational breakdown, or model memory limitations. These terms describe observable symptoms but provide little insight into the structural conditions that produce them. In order to analyse these behaviours more precisely, Cognitive Memoisation adopts the Governance Axes multi-dimensional lens.

The Governance Lens models system behaviour using a set of orthogonal governance dimensions. Each axis represents a distinct property required for stable knowledge interaction. Examples include authority, agency, epistemic custody, constraint enforcement, attention stability, and temporal coherence. Because the axes are independent, pressures can accumulate along one dimension without necessarily affecting others.

Within this model, erosion is interpreted as the result of governance pressure acting on one or more axes during extended interaction. For example, bounded context windows place pressure on the Attention axis because the model cannot maintain awareness of earlier constraints once they fall outside the active context. Context truncation places pressure on the Temporal axis because the reasoning system cannot maintain continuity of knowledge across interaction steps. Similarly, paraphrasing or reinterpretation may place pressure on the Normative Fixity axis when previously defined meanings are no longer applied consistently.

Large Language Model behaviour is frequently discussed in terms such as hallucination, memory limitations, or context length. These descriptions often reflect a category error. They describe symptoms of knowledge erosion rather than the structural conditions that produce it. The Governance Lens allows these behaviours to be analysed in terms of governance pressures acting on systems, including stateless LLM systems.

Once these pressures are visible, engineering responses become possible. Cognitive Memoisation provides such a response by externalising knowledge and enabling a governed knowledge round-trip that restores knowledge stability during interaction.

The Governance Lens therefore provides a diagnostic framework that transforms vague descriptions of LLM behaviour into observable governance conditions. In the context of Cognitive Memoisation, this framework makes it possible to identify where erosion occurs and to design mechanisms that restore knowledge stability during interaction. The lens also supports regression test analysis by providing a consistent method for comparing governance pressure across interaction sequences and experimental runs.

Although introduced here in the context of LLM interaction, the Governance Lens is not limited to AI systems. The framework may also be applied in other domains to analyse governance pressure in systems, institutions, and forms of human behaviour such as fraud analysis. In this sense the lens provides a general method for observing how structural pressures influence behaviour across complex systems.

3. Knowledge Round-Trip Engineering

If knowledge erosion in LLM interaction arises from these identified pressures acting on stateless inference systems, then maintaining knowledge stability requires an engineering response that does not rely on conversational continuity alone. Cognitive Memoisation provides such a response by externalising knowledge into durable artefacts that can be preserved outside the inference system and reintroduced when required. While the Governance Lens provides a useful semantic observation scaffold, Cognitive Memoisation provides the engineering mechanism for restoring knowledge stability.

In conventional interaction with LLM systems, knowledge remains embedded within conversational context. Because that context is bounded and continuously changing, information that was previously established is gradually lost through truncation, paraphrasing, or reinterpretation. Cognitive Memoisation alters this interaction model by capturing knowledge as explicit artefacts rather than leaving it solely within transient inference context.

Once captured, these artefacts may be preserved independently of the model and projected back into inference when needed. In this way, knowledge that would otherwise be lost through context eviction or semantic drift can be restored in a consistent and controlled manner. The inference system therefore operates with a stable reference to previously established information rather than repeatedly reconstructing it during each interaction sequence.

This process establishes a governed knowledge round-trip between human cognition, durable knowledge artefacts, and LLM inference. Knowledge is articulated and captured, preserved outside the inference system, and then reintroduced to support reasoning continuity. Because the preserved artefacts remain external to the model, they are not subject to the same forms of erosion that affect conversational context.

The result is a stable interaction structure in which knowledge can persist across sessions and over time. Instead of repeatedly reconstructing definitions, constraints, and previously established facts, these artefacts can be restored directly when required. Cognitive Memoisation therefore transforms interaction with LLM systems from a fragile conversational process into a governed knowledge engineering process.

4. Epistemic Objects and Memoisation

Cognitive Memoisation implements knowledge round-trip through the use of explicit epistemic artefacts known as Epistemic Objects (EO). An Epistemic Object represents a captured unit of knowledge that has been articulated by a human participant and externalised in a durable form. Once created, an EO can be preserved independently of any single interaction and later projected back into inference to restore knowledge continuity.

Epistemic Objects are designed to be stable and identifiable. The content body of an EO represents the articulated knowledge itself and is treated as invariant once established. If the knowledge represented by an EO changes, a new EO is created rather than modifying the existing one. In this way the system preserves the historical integrity of previously articulated knowledge while allowing new knowledge to evolve.

EOs are organised and referenced through associated artefacts that support their use in reasoning contexts. These structures allow groups of related knowledge objects to be introduced into inference as coherent sets rather than as isolated fragments. The resulting collections provide a stable reference structure that supports reasoning continuity across interaction sequences.

Within Cognitive Memoisation these artefacts form the basis of a memoisation process in which knowledge that has already been articulated and stabilised does not need to be repeatedly reconstructed during later interaction. Instead, previously established knowledge can be restored directly through projection of the corresponding epistemic artefacts.

By representing knowledge as explicit artefacts rather than transient conversational context, Cognitive Memoisation allows reasoning systems to operate with durable references to prior knowledge. This approach enables knowledge to persist across sessions and over time while remaining subject to human governance and correction.

The Epistemic Object model therefore provides the structural foundation that allows knowledge round-trip to be implemented in practice.

5. Cognitive Memoisation Artefact Architecture

Cognitive Memoisation implements knowledge round-trip through a structured set of epistemic artefacts that represent captured knowledge and its organisation within reasoning contexts. These artefacts form the operational architecture used to preserve knowledge outside transient model context and to project it back into inference when required.

The core artefact in this architecture is the Epistemic Object (EO). An EO represents a captured unit of articulated knowledge externalised by a human participant. The body of an EO contains the expressed knowledge itself and is treated as invariant once established. If the knowledge changes, a new EO is created rather than modifying the existing one. This preserves the historical integrity of previously articulated knowledge while allowing new knowledge to emerge.

EOs are referenced and organised through associated artefacts that enable their use within reasoning systems. Epistemic Attributes (EA) provide structured descriptive properties associated with an EO. These attributes may express relationships, metadata, contextual identifiers, or structural information that supports the interpretation and organisation of epistemic artefacts without altering the EO itself.

Reference Objects (RO) provide mechanisms for grouping and organising related epistemic artefacts so that coherent sets of knowledge can be introduced into inference as stable collections. Through ROs, multiple epistemic artefacts may be assembled into structured groupings that support reasoning continuity during interaction.

Together these artefacts form a structured knowledge representation triad consisting of Epistemic Objects, Epistemic Attributes, and Reference Objects. This triad provides the structural basis for organising, relating, and restoring knowledge during interaction.

Within Cognitive Memoisation these artefacts support a memoisation process in which knowledge that has already been articulated and stabilised does not need to be reconstructed during later interactions. Instead, previously established knowledge can be restored by projecting the relevant epistemic artefacts into the inference context.

By externalising knowledge in this way, Cognitive Memoisation allows reasoning systems to operate with durable references to prior knowledge. The resulting interaction model enables knowledge to persist across sessions and over time while remaining under explicit human governance and correction.

The CM-2 protocol defines the core epistemic artefacts, while the normative architecture demonstrates one structural organisation of these artefacts within a working system.

6. Projection and Restoration of Knowledge

Cognitive Memoisation restores knowledge stability during interaction by projecting preserved epistemic artefacts back into inference when required. Because Epistemic Objects exist outside the transient conversational context of the model, they remain available even when earlier interaction history has been truncated or replaced.

During extended interaction with stateless LLM systems, previously established information may fall outside the active context window. When this occurs the model can no longer rely on that information during reasoning. Cognitive Memoisation addresses this condition by reintroducing the relevant epistemic artefacts into inference so that previously established knowledge can be restored.

Projection occurs when selected epistemic artefacts are introduced into the active reasoning context of the model. These artefacts provide stable references to definitions, constraints, or previously articulated knowledge that would otherwise need to be reconstructed through conversation. Because the artefacts represent externalised knowledge rather than conversational fragments, they provide a consistent basis for reasoning continuity.

In practice, epistemic artefacts are often introduced into inference as structured groupings rather than as isolated objects. These collections allow related knowledge elements to be restored together, providing a coherent knowledge context for the reasoning system. By restoring groups of related artefacts, Cognitive Memoisation reduces the risk of partial knowledge reconstruction that can otherwise lead to semantic drift or constraint erosion.

The projection of preserved artefacts therefore provides a mechanism for restoring knowledge continuity during interaction. Instead of relying solely on the persistence of conversational context, the reasoning system can re-establish previously articulated knowledge directly from preserved epistemic artefacts.

Through this mechanism Cognitive Memoisation enables knowledge to persist across interaction sequences, sessions, and time. Even when the conversational context of the model changes, the preserved artefacts remain available for reintroduction into inference, allowing reasoning processes to resume with stable references to prior knowledge.

7. Normative Architecture and Client Memoisation

The preceding sections describe the Cognitive Memoisation architecture as it has been demonstrated through the CM-2 Normative Architecture. This architecture provides a working reference implementation in which epistemic artefacts are preserved and projected into inference to maintain knowledge continuity during interaction.

In this reference architecture, epistemic artefacts may be preserved within a durable substrate associated with the reasoning system. These preserved artefacts allow previously articulated knowledge to be restored when conversational context is lost.

However, the CM-2 protocol itself does not require that epistemic artefacts be permanently stored by a platform. Cognitive Memoisation is designed so that epistemic artefacts may also be maintained by the human participant through client-side memoisation. In this model, Epistemic Objects captured during interaction may be preserved locally by the client and supplied to the inference system when required.

Client memoisation extends the Cognitive Memoisation model by ensuring that humans retain sovereignty over their knowledge artefacts. A platform may choose to store Epistemic Objects within its own durable substrate, but it is not required to do so. Because the epistemic artefacts remain portable, the inference system can obtain the required knowledge objects directly from the client when they are projected into reasoning contexts.

This design ensures that Cognitive Memoisation remains independent of any particular platform implementation. Knowledge artefacts remain under human control and may be preserved, transported, and reintroduced across different reasoning systems. In this way the CM-2 protocol supports distributed cognition while preserving the sovereignty of human participants over their own articulated knowledge.

CM-2 does not depend on one architecture, the CM-2 Normative Architecture is just one effective realisation.

8. Mechanical Extraction of Thought

Cognitive Memoisation relies on the capture of articulated knowledge in the form of Epistemic Objects. These artefacts originate from human reasoning and must therefore be extracted from the thought processes of participants during interaction.

The process by which articulated knowledge is transformed into structured epistemic artefacts is referred to as Mechanical Extraction of Thought. Through this process, statements, definitions, constraints, and conceptual structures expressed during interaction are identified and externalised as Epistemic Objects that can participate in the Cognitive Memoisation architecture.

Mechanical Extraction of Thought does not require the model itself to generate knowledge artefacts automatically. Instead, the process may be performed by human participants, by supporting tools, or by automated systems that assist in identifying candidate knowledge structures during interaction.

Once extracted, these epistemic artefacts may be preserved and organised within the Cognitive Memoisation framework and subsequently projected back into inference to maintain knowledge continuity.

Mechanical Extraction of Thought therefore provides the complementary process that enables Cognitive Memoisation to operate in practice. While Cognitive Memoisation governs the preservation and restoration of knowledge artefacts, Mechanical Extraction of Thought governs the capture of those artefacts from human reasoning processes.

A more detailed description of Mechanical Extraction of Thought is provided in the companion paper dedicated to this topic.

9. Cognitive Memoisation as Open Knowledge Infrastructure

Cognitive Memoisation introduces a structured approach to preserving and restoring knowledge during interaction with stateless inference systems. By externalising articulated knowledge as Epistemic Objects and enabling their controlled projection into inference, Cognitive Memoisation transforms interaction with LLM systems from a fragile conversational process into a governed knowledge engineering process.

The CM-2 protocol provides a portable mechanism for representing and exchanging epistemic artefacts across reasoning systems. Because these artefacts remain external to any single platform implementation, knowledge captured through Cognitive Memoisation can be preserved, transported, and reintroduced across different inference environments. This portability allows knowledge engineering work performed with one system to remain usable within another.

An important design principle of Cognitive Memoisation is that human participants retain sovereignty over their knowledge artefacts. Platforms may choose to provide durable storage and management of Epistemic Objects, but the protocol does not require this. Knowledge artefacts may instead be preserved through client memoisation and supplied to inference systems when required. This design ensures that knowledge remains under the control of the human participant rather than the platform.

Cognitive Memoisation is therefore intended to function as open knowledge infrastructure for human-AI collaboration. The protocol is released under an open licence to encourage adoption across vendor platforms and to support interoperable knowledge engineering practices.

Open infrastructure requires clear governance in order to preserve the integrity and continuity of shared knowledge systems. For this reason the Cognitive Memoisation corpus and protocol development are supported by a Public Stewardship Model. This model establishes a framework for maintaining the integrity, corrigibility, and long-term continuity of the Cognitive Memoisation body of work while allowing open participation and adoption.

The conceptual model presented in this paper is supported by a set of normative documents that define the CM-2 protocol, the Cognitive Memoisation normative architecture, and the reference object collection bootstrap used to initialise Cognitive Memoisation systems. These documents provide the formal definitions and implementation guidance required to realise the Cognitive Memoisation framework in operational environments.

Conclusion

Cognitive Memoisation provides a method for preserving human knowledge during interaction with stateless inference systems. By externalising articulated knowledge as durable epistemic artefacts and enabling their governed restoration into inference, Cognitive Memoisation allows knowledge to persist across sessions and over time. In doing so, it establishes a foundation for durable, portable, and openly governed knowledge engineering in human-AI collaboration.