When Meaning Becomes Architecture: Knowledge Bases, Self-Models, and Conceptual Mediation in Agentic Systems

A knowledge base is commonly described as a store of facts. An agent asks it a question, retrieves what is relevant, and uses that material to produce an answer or choose an action. As a first approximation this is useful, but it hides most of what matters. A knowledge base never contains facts without form. Its contents have been divided into entities, properties, classes, relations, events, documents, or vectors. Someone, or some process, has decided what counts as the same object, which distinctions matter, and how one item relates to another. Even an apparently unstructured pile of documents is surrounded by decisions about chunking, metadata, similarity, ranking, provenance, and access.the ontology is the hidden architecture

The knowledge base is therefore not merely what an agent knows. It is part of the means by which anything becomes knowable to that agent at all. The earlier page in this series argued that words are tools that mediate human activity. This page asks what changes when a system can store, retrieve, combine, and act through such tools, and when its distinctions are no longer only discussed but implemented as persistent, executable parts of the system.

The central proposition is this: a knowledge base is part of an agent's mediating apparatus. It is a structured set of conceptual tools through which the agent identifies objects, interprets situations, selects actions, and models itself.

1. From Communicative Tools to Cognitive Architecture FoundationalKnowledge that endures for decades — core principles

ForFor a human community, language supplies shared distinctions. For an artificial agent, the same work is done by a range of technically different things:

• prompts and system instructions;
• schemas and controlled vocabularies;
• ontologies and knowledge graphs;
• document stores and vector embeddings;
• episodic records of past interactions;
• tool descriptions, policies, and plans;
• user models and self-models.

These are different in implementation, but each can take part in mediating activity. A caution is needed about what this claim supports. The system does not necessarily understand these structures in the human, experiential sense. The weaker and more defensible claim is functional: the structures change what the system can discriminate, retrieve, infer, communicate, and do. That is the level at which this page makes its arguments, and it is the same level at which the consciousness-analogue page in the Modelling the Self series proposes its self-model: as an architectural analogue, not as a claim that the architecture is conscious.functional, not phenomenal cf. consciousness-analogue

2. Knowledge Bases Contain Claims, Not Reality Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA conventional representation shows the problem clearly:

Author ──writes──> Article
Article ──hasTopic──> Activity Theory
Activity Theory ──concerns──> Mediated Activity

This looks like a set of plain facts. Every element, however, depends on a modelling choice. Is the author represented as a person, an account, a lecturer, or a legal identity? Is writing an event, a relation, a role, or a process? Is an article a document, a publication, a web resource, an intellectual work, or all four at once? Is Activity Theory a topic, a theory, a research tradition, or a named entity? Does hasTopic record the author's intention, the text's occurrence of a theme, or a reader's interpretation?

A knowledge base cannot include everything. It has to reduce the world to selected differences that its designers or users consider consequential. That leads to a compact formulation: a knowledge base stores not reality but claims expressed through a representational scheme. Fuller still, it is a historically situated collection of claims whose possible form is constrained by the representation system that holds them.the map is not the territory

3. Four Kinds of Knowledge Structure Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AgentAgent architectures use several arrangements, and it helps to distinguish at least four.

Document collection

Stores human-produced text or media. A query retrieves passages, which are supplied as context. Meaning remains largely implicit in the documents and has to be reconstructed during retrieval and generation.

Vector store

Represents items as positions in a high-dimensional space built from learned patterns. A query is embedded, compared by similarity, and the nearest items are ranked. Relations are based on statistical proximity rather than being explicitly named.

Knowledge graph

Represents identifiable entities and explicit relations between them. Knowledge graphs are graph-based data models of this kind, studied extensively as a formal topic[4]. They support traversal, provenance, and connections that do not depend on textual similarity.

Formal ontology

Defines permitted classes, relations, constraints, and sometimes inference rules. A rule might state that every research article is a document, that every document has at least one creator, and that no person is a document.

Real systems often combine these. An agent might use document retrieval for detail, vector search for semantic proximity, a knowledge graph for explicit relationships, and an ontology to validate claims and support inference. Retrieval-augmented generation, the pattern in which a language model is supplied with material retrieved from such stores, combines parametric memory held in the model's weights with non-parametric memory held in an external index[1]. The retrieval-augmented generation page covers the pipeline itself. This page is concerned with what the pipeline does to meaning.

4. Retrieval Is an Interpretive Act Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

RetrievalRetrieval is often presented as though the agent simply looks up the answer. In practice it requires the system to construct or receive a query, rank candidate material, and decide what goes into the context window. That means every retrieval involves at least a representation of the present problem, assumptions about relevance, a mechanism for comparing the problem with stored material, a threshold for inclusion, an ordering of what was found, and decisions about what is left out.like a search query shaping what you find

Text description of the diagram

Six boxes in a vertical chain: initial query; provisional interpretation; retrieval operation; selected context; revised interpretation; response or action. Downward arrows link them in order. A dashed return arrow runs from the revised interpretation back to the retrieval operation, labelled "interpretation shapes retrieval".

Context, then, is not simply added after interpretation. A provisional interpretation determines what is retrieved, and retrieval then reshapes the interpretation. The circularity is the important part: an agent retrieves according to what it currently takes the situation to mean, and what it retrieves changes what the situation can mean to it. This is the point where the account of meaning as interpretation, from the companion page, becomes an operational pipeline.

5. The Resources Mediate the Task Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe mediating structure of an agentic system can be drawn as a triangle of the kind used in activity theory, with the system's resources at the apex and the task object at the base:

            SYSTEM RESOURCES
  prompts, schemas, tools, memories,
  ontology, policies, knowledge bases
              /           \
             /             \
        AGENT ----------> TASK OBJECT

The resources do more than make a predefined action cheaper to carry out. They shape how the task is represented, what counts as relevant evidence, which actions are recognised as available, what outcome counts as success, which constraints are visible, and which consequences can be anticipated. A calendar API, for example, does not simply add one action to an otherwise unchanged conversational model. Its schema brings in entities such as meetings, attendees, organisers, locations, time zones, recurrence rules, and permissions, and those distinctions become part of the system's actionable world. The agent's ontology is, in a real sense, distributed across its tools.the tool's schema redefines the ontology

6. Tool Schemas as Micro-Ontologies Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

EveryEvery tool made available to an agent presents a compact account of a possible world. Consider a simple schema:

{
  "action": "create_meeting",
  "participants": [],
  "start_time": "",
  "end_time": "",
  "location": "",
  "agenda": ""
}

The schema implies that meetings are creatable objects, that meetings have participants, that time can be represented as a bounded interval, that a meeting may have a location and an agenda, and that the acting system can initiate the relevant process. What it leaves out matters just as much: the authority to invite particular people, accessibility requirements, conflicting obligations, informal power relations, whether attendance is optional, and whether the meeting should happen at all.

A tool description is therefore more than an instruction for calling software. It is a small operational ontology that specifies which entities, properties, and transformations the agent can act on. Forms, API specifications, database schemas, menus, workflow states, function signatures, permissions, and error codes all work the same way. Each turns a selection of distinctions into operational possibilities.the map is not the territory

7. From Language to Consequence Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe distinction between a language model and the wider agentic system supporting it is central here. The model proposes continuations of text. The architecture determines which of those continuations become consequential. A model may generate the sentence "I will schedule the meeting for Tuesday." Without connected tools that is only generated language. In an agentic system the same output may be converted into a call such as create_calendar_event(...), but only after several things are in place: an interpretation of the user's intention, a model of the available tool, structured arguments, permissions, validation, state tracking, and an account of success or failure.

At that point meaning has become operational. The word "schedule" is connected to a change in a shared environment. This does not show that the system experiences meaning. It shows that meaning-bearing signs can take part in causal, rule-governed activity. That is a different claim, and the two should be kept apart:

Semantic efficacy is not the same claim as phenomenal understanding. An agent's representations can have functional consequences without establishing that the system has subjective experience.cf. the old symbol grounding problem

8. Knowledge About the World and About the Agent Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA conventional knowledge base models objects in an external domain: customers, projects, documents, deadlines, meetings. A self-modelling architecture adds the system itself as one of the objects:

Agent
  ├── hasCapability
  ├── hasPermission
  ├── hasGoal
  ├── hasUncertainty
  ├── hasMemory
  ├── hasCurrentPlan
  └── hasPriorAction

This opens two broad possibilities. The first is the self-model as an opaque object. The system represents itself explicitly, with claims such as "Agent-1 hasCapability WebSearch", "Agent-1 lacksPermission DeleteRecord", "Agent-1 hasConfidence 0.62", and "Agent-1 isExecuting Plan-7". These claims can be inspected, questioned, revised, and used in planning, so an agent might reason that a task needs a resource it cannot reach and therefore should ask the user for it. Here the self-model is knowledge about the system.

The second is the self-model as transparent mediation. Some representations shape processing without being available as inspectable objects. Relevance thresholds, trust scores, habitual routing, default priorities, hidden system instructions, learned associations, and salience mechanisms are examples. They do not merely tell the agent what it is. They help determine how the agent encounters a situation in the first place.

The Modelling the Self series draws this same distinction between an opaque self-model, which the reasoning process can consult, and a transparent one, whose effects enter perception without appearing as a model at all. The consciousness-analogue page takes up what that distinction implies for theories of consciousness, and it stops short of claiming that either version is conscious.cf. one front end, four interchangeable last stages

9. The Self-Model Is Also a Tool FoundationalKnowledge that endures for decades — core principles

AA self-model is often imagined as a static description: "I am Agent X. I have tools A, B, and C. My current task is Y." Within an activity-oriented account it is better treated as a mediating artefact. It helps the system estimate what it can do, recognise when it needs help, predict the consequences of its actions, distinguish its own actions from external events, keep commitments over time, notice contradictions between goals, explain or justify earlier actions, and revise its plan after failure.

So the agent does not merely possess a self-model. It acts through that model, and the action changes the model in turn:

Self-model
    │
    ▼
Action selection
    │
    ▼
Outcome and feedback
    │
    ▼
Updated capability, history, or confidence
    │
    └──────────────> Revised self-model

A failed tool call may lower the system's estimate of its own capabilities. A user correction may change its model of the task. A completed action enters its episodic history. A newly granted permission widens the set of actions it treats as available. The loop runs both outward, from the agent into the world, and inward, from the agent back into its own model of itself.

10. Identity Across Time Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA self-model raises a question that connects ontology with memory: what makes the agent at one time the same agent as at another? Possible answers appeal to continuity of process, identifier, conversation, memory, goals, policy, software instance, ownership, commitments, or narrative. These criteria can conflict. The underlying language model may be upgraded while the conversation history remains. Some memories may be deleted. Tool permissions may change, the public name may stay the same, and the instructions may be replaced. Is it still the same agent?

This is not decoration for a philosophy seminar. The answer affects accountability, audit trails, delegation, permissions, trust, promises, the attribution of earlier actions, and the interpretation of stored memories. A robust agent architecture therefore needs an operational account of identity, even if it avoids any grand claim about personhood. The account it needs is the one that says which continuity conditions the system's records actually depend on.

11. Memory Is Not One Thing Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AgentAgent memory should not be treated as a single store. Psychology already distinguishes memory systems: Tulving separated episodic memory, which concerns unique, personally situated experiences, from semantic memory, which holds general facts and meanings shared with others[3]. Language-agent designs have borrowed and extended that kind of distinction. The CoALA framework organises such designs around memory modules and action spaces, drawing on the older cognitive-architecture tradition[2]. A useful functional breakdown for an agent looks like this:

Working context

Material currently available to generation or reasoning.

Episodic memory

Records of particular interactions and events, such as "on 4 October the user requested a summary of a particular document."

Semantic memory

More general claims abstracted from particular events, such as "the user prefers concise summaries."

Procedural memory

Patterns that govern how tasks are performed, such as "when reviewing a source, check its provenance before summarising it."

Social or institutional memory

Shared records that exist beyond any one agent: policies, project logs, team decisions, version histories.

Self-model

Claims about the system's own identity, state, capabilities, limits, and commitments.

These stores can disagree. A current request may conflict with a saved preference. A procedural rule may conflict with a task goal. A self-model may claim access to a tool that is no longer available. An architecture with this structure has to negotiate contradictions between stores, not simply retrieve from them.

12. Contradictions as Drivers of Development Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

IfIf tools and activity change one another, the knowledge structures an agent uses should not be treated as permanently settled. The companion page on words as tools describes how contradictions in an activity system can drive its development. The same pattern appears inside an agent. Contradictions may arise between the ontology and a newly encountered case, between the self-model and actual performance, between the user model and a new instruction, between two sources making different claims, between a policy and a goal, between a plan and the tools permitted for it, or between a stored classification and a changed external reality.

A simple system might suppress one side of such a contradiction. A more capable one could represent it:

Source A asserts P.
Source B asserts not-P.
The claims apply to different dates.
Source A is authoritative for policy.
Source B describes observed practice.
Further resolution is required.

This suggests a view of intelligence in which the capacity to recognise, locate, and productively respond to contradictions may matter as much as possessing a perfectly consistent knowledge base. It also offers a computational analogue of concept development. When an object fails to fit an existing schema, the system can reclassify it, create a new category, subdivide an old one, revise a rule, qualify a claim by context or time, or accept that the plurality is unresolved.

13. The Politics and Ethics of Agent Knowledge Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

IfIf a knowledge base mediates action, its classifications have consequences. Several questions follow. Who supplied the original categories? Whose language is treated as authoritative? Which objects can be represented, and which relationships stay invisible? Can people inspect and contest the stored claims made about them? Does the system distinguish assertion from verified fact? Are historical claims given temporal boundaries? Can a source withdraw or revise what it has supplied? Does the agent preserve provenance? Can the ontology represent uncertainty and disagreement? And at what point does a model of a user become an imposed identity?

The last question matters most for self- and user-modelling. A model built to support helpful adaptation can harden into a constraint. A stored note such as "user prefers concise responses" can, through repeated application, turn into "always omit extended explanation for this user." A preference inferred from history has then become an active rule that the user may never see or be able to change. The same process happens socially when descriptive categories move into institutional systems.

14. A Unifying Framework Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe argument can be summarised as five levels, each with the question it answers:

Five levels of an agentic knowledge architecture

Level	Contents	Question it answers
1. Signs	Words, symbols, identifiers, embeddings, interface elements	What distinctions are available?
2. Representations	Claims, documents, triples, records, schemas, models	How are things represented?
3. Knowledge organisation	Ontologies, graphs, indexes, provenance, temporal scope, validation rules	What relationships and inferences are permitted?
4. Activity	Retrieval, planning, tool use, communication, intervention, collaboration	What can the system do through these representations?
5. Reflexivity	Memory, performance monitoring, user modelling, self-modelling	Can the system represent and revise its own role in the activity?

The levels are recursive. Signs structure representations. Representations organise activity. Activity produces new signs and records. Records revise knowledge. Knowledge changes future activity. Reflexive activity changes the agent's model of itself, and that model then governs what kinds of activity it takes up.

15. Conclusion FoundationalKnowledge that endures for decades — core principles

AA knowledge base is often treated as a warehouse into which facts are placed and from which answers are retrieved. For an agentic system that picture is inadequate. The structure of the knowledge base helps determine what the agent can recognise, which relations it can follow, which questions it can formulate, and which actions it can perform. Words become categories, categories become schemas, schemas become knowledge structures, and knowledge structures become conditions of action. Once they are connected to planning and tools, these representations no longer only describe a world. They take part in changing it.

The same holds reflexively. When an agent represents its own capabilities, history, uncertainty, goals, and limits, it makes itself one of the objects in its actionable world. That self-model mediates future activity, and the results of that activity feed back into the model.

None of this establishes consciousness or human-like understanding. It supports a narrower but still important claim: meaning-bearing structures can become causally and organisationally significant within an artificial activity system. They can mediate the relationship between an agent, its environment, other participants, and its own continuing model of itself. Knowledge becomes agentic when representations do not merely inform outputs but take part in selecting, constraining, and revising action. Meaning, on this account, is not simply found in words, databases, or models. It is enacted through the changing relationships among representation, retrieval, interpretation, and action.

Continue With Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

• LLM Orchestration, Context Engineering & Agentic AI — the technical stack for chaining model calls, managing context, and running agentic loops.
• Retrieval-Augmented Generation — the retrieval pipeline and its limitations, which section 4 reads as an interpretive act.
• Does a Self-Model Get You Anywhere Near an Analogue of Consciousness? — what the opaque and transparent self-model distinctions do and do not imply.

References

1. P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. Yih, T. Rocktäschel, S. Riedel, and D. Kiela, "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks," Advances in Neural Information Processing Systems 33 (NeurIPS 2020), pp. 9459–9474. https://arxiv.org/abs/2005.11401
2. T. Sumers, S. Yao, K. Narasimhan, and T. Griffiths, "Cognitive Architectures for Language Agents," arXiv:2309.02427, 2023 (revised 2024). https://arxiv.org/abs/2309.02427
3. E. Tulving, "Episodic and Semantic Memory," in Organization of Memory, eds. E. Tulving and W. Donaldson, Academic Press, 1972, pp. 381–403.
4. A. Hogan, E. Blomqvist, M. Cochez, et al., "Knowledge Graphs," ACM Computing Surveys 54(4), 2021, article 71. https://doi.org/10.1145/3447772