Learning as the Assembly of Temporary, Extended Knowledge Systems: Connectivism With AI, Documents and Knowledge Graphs Added

Pedagogy,Pedagogy, Andragogy, Heutagogy, Rhizome asks who decides what gets learned, and shows control moving from institution to instructor to learner to community. Learning as a Network adds a second axis: knowledge and learning as properties of a network, not just contents placed in one head. Bring AI, persistent documents, and knowledge graphs into that network and a further question appears, one neither page fully answers on its own. It is no longer only who decides, or even what the network looks like. It is where the network extends, how long it lasts, what should survive once it dissolves, and how its components earn the learner's trust.

The Network Has No Natural Edge

SiemensSiemens and Downes already establish that knowledge is distributed across people, artefacts, and systems, and that learning is the growth and traversal of connections rather than the transfer of content into a mind. Pushed further, that network has no single point at which it naturally starts or stops. It runs inward through neural organisation and the nervous system, outward through bodily action and what gets loosely called muscle memory, out again into notes, diagrams, and other representations, across immediate relationships and wider personal learning networks, and now into search engines, knowledge graphs, and AI systems. To count as part of a learner's network, a component has to make a material difference to their ability to perceive, act, interpret, judge, adapt, or reconnect. Some components — a well-practised sensorimotor pattern — stay tightly coupled for years, while others, a colleague consulted once or an AI conversation opened to test one idea, participate for minutes.

Treating the boundary as functional and temporal rather than fixed is this page's own extension of Siemens's and Downes's picture, not a claim either of them makes directly. It reframes the learner: not someone who simply has a network, but someone who repeatedly assembles, inhabits, governs, and dissolves one.extends connectivism across time, not Siemens's own claim

A Network Can Be Temporary Even When Its Effects Aren't

AA learning network can exist for years or for the duration of one problem. Faced with an unfamiliar situation, someone draws on foundational knowledge, consults a colleague, searches the literature, asks an AI system to generate candidate explanations, tests those explanations against sources or observation, records a conclusion, updates a knowledge graph, and then lets that particular configuration dissolve. The network was temporary. What it leaves behind need not be: a durable change in understanding, an embodied capability, a written record, a new relationship between concepts, a reusable route back to an authoritative source, a sharper sense of where the uncertainty sits, or simply a resolved problem and nothing more than that. The reverse failure mode matters just as much: a repository, an AI assistant, or a community can stay continuously available while changing almost nothing about what a person can actually understand or do. Persistence of access is not the same thing as learning having happened.

Three Kinds of Persistence

ThatThat distinction splits three ways, because the three vary independently of each other:

Persistence of the network
        -- do the people, tools, communities and services stay accessible?
Persistence of the representation
        -- do the notes, documents, citations and graph relationships stay recoverable?
Persistence of capability
        -- can the learner still act, explain, judge or reconstruct the solution once the network is gone?

A brief conversation can create durable understanding that outlives it completely. Fully accessible and permanent, an archive can still remain unintelligible to anyone who didn't build it. A short-lived project group can leave behind thoroughly reusable documents, while a long-standing community can depend almost entirely on tacit knowledge nobody wrote down. Knowing which of the three a given learning episode actually produced is more useful than asking whether "the resource is still there."

What Kind of Material Is This?

NotNot all material deserves the same relationship with the learner. Five categories cover most of what gets called learning:

• Foundational material changes slowly and enables everything built on top of it: core concepts, disciplinary vocabulary, causal models, characteristic methods, and the field's central ethical and epistemic principles. This is the material that usually warrants durable internalisation, because a learner needs enough of it on hand to recognise a problem and spot an implausible answer without stopping to consult the network first.
• Organisational material is the map of a domain rather than its content: which concepts relate to which, which disagreements actually define the field, where authoritative material lives, and which pathways tend to answer a given kind of question. The learner doesn't need to retain every item the map points to, only an accurate enough map and the ability to reconnect to what it points at.
• Procedural material concerns how to perform, investigate, or evaluate. Some of it becomes embodied through repetition. What gets called muscle memory is not memory stored in the muscle; it is learned sensorimotor organisation built through neural plasticity, perception, action, and feedback. Other procedures stay appropriately distributed across checklists, instruments, software, and collaborators, and the question that matters is not whether every step has been memorised but whether the resulting performance is reliable, adaptable, and properly governed.
• Situational material belongs to one case: one learner's circumstances, one project's constraints, the state of one system, the assumptions behind one analysis. It needs retaining only long enough to support the decision and its accountability, not long enough to become part of permanent personal knowledge.
• Ephemeral material is valuable mainly because it is current: system status, organisational roles, live regulation, software behaviour. Memorising it is usually the wrong investment. Knowing that it is time-sensitive, where the current authoritative version lives, how to check its date, and what would make it expire is worth more than the content itself. Forgetting obsolete material here is not a failure of memory; it is the correct response, provided the capacity to retrieve and evaluate its replacement survives the forgetting.

The Different Senses of AI "Memory"

UsingUsing an AI system inside a learning network needs distinctions that everyday talk about "AI memory" usually blurs together.

Model parameters hold statistical regularities acquired during training. They are not a searchable record of individual sources, and they are not a reliable log of a learner's own previous conversations. From those patterns, a model can generate a useful explanation without being able to say where, exactly, any one part of it came from or whether it is still current.

The context window is everything available to the model while it produces its current response: system instructions, recent messages, retrieved documents, tool results, and the answer being generated. Anthropic's own documentation describes the context window as "a working memory for the model," distinct from the much larger corpus it was trained on¹. Taken literally, that makes it an active workspace rather than a durable record: material leaves it when a conversation ends, the window fills, older content is dropped, or the conversation is compacted into a summary.working memory, not a store

Application-level memory is a separate thing again: information stored outside the active context and brought back in during a later session. Anthropic's memory tool, for instance, works through a directory of files that the surrounding application lets the model create, read, update, and delete between conversations². In Claude Code, persistent instruction files and automatically accumulated notes work the same way: loaded into a new session as context, not as an unchangeable rule baked into the model. The model does not simply remember: an application around it stores a representation, chooses what to retrieve, and places that back into context at the point it seems relevant. That is a real convenience, and it is also a knowledge-management process with all the usual knowledge-management questions attached: who chose what to keep, whether it is still accurate, what contradicting evidence has turned up since, whether the learner can inspect and correct it, and when it should be retired.

Long-term written documents differ from both, because they can be stable, inspectable, and governed independently of any one conversation or application: a document that preserves claims and their supporting evidence, provenance and dates, unresolved disagreements, decisions and the reasoning behind them, and explicit uncertainty can be checked against primary sources, revised on purpose, versioned, shared, and used by different tools later. Conversational fluency creates a strong illusion of continuity; a conversation that reads as coherent is not the same thing as a record anyone could audit afterwards. The practical rule that follows is to treat the conversation as a working session and promote what matters out of it into a properly structured, governed document, rather than letting the chat log double as the record of what was learned.

Knowledge Graphs as Connective Tissue

DocumentsDocuments are good at preserving an extended argument with its context intact, and correspondingly bad at making relationships across several documents easy to see. A knowledge graph complements a document by making selected entities, claims, and relationships explicit and machine-traversable: a concept and its prerequisites, a claim and the evidence for or against it, a method and the conditions under which it applies, a decision and the assumptions it depended on. The schema behind that graph, the set of entity and relation types it allows, is itself a formal ontology in the computer-science sense, and this site's own treatment of that sense is explicit that a formal ontology needs to carry its scope, provenance, version, and temporal validity alongside the structure itself, not bolted on once something has already gone wrong. Being expressed in graph form doesn't make any of that automatically true on its own. A graph is a set of assertions in structured clothing, and its edges need the same things its schema does: provenance, scope, and, where it matters, a confidence level. "Source A supports Claim B under Condition C as of Date D" is a safer thing to record than the bare "A proves B," because the first version carries its own temporal validity on its face and tells a later reader exactly what would make it stop applying.

A 2024 roadmap on unifying large language models with knowledge graphs describes the relationship between the two as reciprocal rather than one feeding the other: language models can help extract candidate entities and relationships from text, while graphs supply the structured, checkable grounding that later model interactions can draw on³. According to the same source, automated graph construction is still vulnerable to extraction errors, inconsistent naming, and the kind of error that compounds once it propagates into a structure other steps rely on. The workflow that respects that limitation is not "the model reads documents and the model's output becomes the graph's truth." It is closer to: the model proposes structure, a human or a community validates it, the graph records the relationship together with its provenance, later contexts retrieve what's relevant, and new evidence revises both the graph and the documents it was drawn from.

Communities of Practice as Active Knowledge Infrastructure

AA community of practice actively shapes what currently counts as competent practice, which problems are treated as important, which sources are trusted, how evidence gets interpreted, what stays disputed, and when an established answer has stopped fitting current circumstances. That makes a community especially important for exactly the kinds of knowledge least suited to a static document: organisational, procedural, situational, and ephemeral material, including tacit distinctions nobody has written down yet and the early sense that a formal document has fallen behind what practitioners are actually doing.

None of that makes a community automatically correct: the same mechanisms that let it preserve live tacit knowledge also let it reproduce local habit, unexamined assumption, status hierarchy, obsolete practice, groupthink, and the systematic exclusion of knowledge that doesn't come from inside the group. AI can help a community marshal its own material: finding related documents, summarising discussions, flagging recurring questions, proposing candidate graph relationships, pointing out apparent contradictions, and drafting a provisional synthesis for people to react to. What it should not do is quietly become the voice of the community. An output that represents shared policy, disciplinary consensus, or contested professional judgement needs a human who can be identified as having reviewed it, especially where that output is going to be relied on by people who weren't in the room.

A Practical Learning and Knowledge Cycle

PutPut the pieces above together and a cycle with roughly ten stages emerges, useful less as a rigid procedure than as a checklist of the steps a learner can skip without noticing:

Text description of the diagram

Ten stages arranged clockwise in a loop: Encounter, Assemble, Explore, Ground, Structure, Test, Consolidate, Share, Revalidate, and Dissolve. An arrow runs from each stage to the next, and the final stage, Dissolve, arrows back round to Encounter, showing the cycle repeating rather than ending.

1. Encounter a question, anomaly, task, or unfamiliar situation.
2. Assemble a temporary network drawn from internal knowledge, practical experience, documents, colleagues, community discussion, data, an AI system, and whatever specialist tools the problem calls for.
3. Explore with the AI generating candidate questions, alternative explanations, counterexamples, or search terms, favouring breadth over premature certainty at this stage specifically.
4. Ground the claims that actually matter against evidence, observation, and authoritative sources, and keep a generated suggestion, a retrieved statement, a personal interpretation, a community judgement, and verified evidence distinct from one another rather than letting them blur into a single impression.
5. Structure the emerging understanding into notes, maps, tables, citations, and candidate knowledge-graph relationships.
6. Test it by applying, explaining, critiquing, or transferring it, with AI useful here for asking hard questions, presenting deliberately broken examples, arguing the opposing position, or asking for the reasoning to be reconstructed rather than restated.
7. Consolidate selectively: internalise the foundational material, practise the procedures, document the situational material, timestamp the ephemeral material, add organisational relationships to the graph, and move the conclusions that matter out of conversational context and into a durable document.
8. Share and negotiate the result with the relevant community, where it can be challenged, refined, adopted, or rejected.
9. Revalidate time-sensitive claims and anything AI-produced the next time it actually gets reused, rather than trusting that last time's check still holds.
10. Prune and dissolve the obsolete links, superseded claims, and working material that served the episode but doesn't need to outlive it, leaving behind only the capabilities, records, and relationships actually worth keeping.

Trust Across the System

TrustTrust attached globally to a person, a document, a community, a graph, or an AI system is the wrong shape for it. Trust needs calibrating against a specific claim, a specific task, and a specific consequence, which means asking the same set of questions of any node regardless of what kind of node it is:

• Competence. Is this node actually capable in the relevant domain?
• Provenance. Where did this claim or relationship come from?
• Currency. When was it valid, and when does it need rechecking?
• Independence. Are apparent confirmations genuinely independent, or one source echoed several times?
• Transparency. Can the reasoning or evidence behind it actually be examined?
• Consequence. What happens if this is wrong?
• Replaceability. Is there another route if this particular node fails?
• Dependency. Is relying on this component building the learner's own capability, or quietly substituting for it?
• Contestability. Can the learner or the community actually challenge and correct it?
• Ownership. Who remains responsible for the judgement that follows from trusting it?

That list extends ordinary source-and-claim scepticism in two directions at once: toward the receiver (is this actually developing the learner's own capability, Dependency, above), and toward time (when was this true, and when does it need checking again, Currency). A source can pass every traditional test of reliability and still fail one of those two.

A Multidimensional Extension of the Learning Theories

Pedagogy,Pedagogy, andragogy, heutagogy, and rhizomatic learning are usually presented as a single ladder, control moving step by step from institution to individual to community. Four intersecting dimensions describe the same territory more completely, and combining them this way is this page's own synthesis, not a claim any one of the cited theorists makes:our own synthesis, not any one theorist's

Governance -- who directs the learning?
        pedagogy: a teacher or institution configures the network
        andragogy: the learner navigates a largely preconfigured network
        heutagogy: the learner determines purposes and configures the network
        rhizomatic learning: participants negotiate the network and its purposes together

Distribution -- where does the capability actually live?
        neural, embodied, material, documentary, interpersonal,
        communal, institutional, graphical, computational

Temporality -- what persists once the episode ends?
        the original network itself, a record of the interaction, a document,
        a graph relationship, a route back to a source, a human capability,
        or just the resolved effect of one temporary episode

Epistemic role -- what should the learner's relationship to this material be?
        internalise it, embody it through practice, understand its organising
        principles, preserve it externally, know how to retrieve it,
        revalidate it on every use, or deliberately let it expire

Connectivism supplies the network itself; heutagogy supplies self-determination, and rhizomatic learning, collective negotiation. Embodied and distributed accounts of cognition locate real learning in neural, bodily, and material activity, not only in propositions, while communities of practice supply the live professional interpretation no document fully captures. Documents, in turn, supply durable, inspectable records; knowledge graphs make selected relationships explicit and traversable; and AI supplies a powerful but fallible mechanism for exploring, transforming, retrieving, and provisionally synthesising all of the above.

The position that falls out of combining them might be called temporal, embodied, epistemic connectivism: learning as a change in the capacity to assemble, use, evaluate, document, reconfigure, and appropriately dissolve neural, bodily, social, documentary, and computational networks over time. Maturity, on this account, is knowing what must be internalised, what should be practised until it is embodied, what can safely stay distributed, what needs a durable written record, what belongs in a graph, what an AI can usefully help explore, what a community has to negotiate rather than settle alone, what needs independent verification, what must be revalidated at every use, and what should simply be allowed to disappear.

That leaves the original question answered by neither of the two extremes it's usually framed as. Not "I decide everything I learn," and not "the community is the curriculum" on its own. Closer to this: a learner participates in constructing, testing, and governing the temporary and persistent networks they learn through, preserves what those networks produce in forms suited to its purpose, and stays responsible for deciding what the network is allowed to make them believe and able to do.

Related Topics

• Pedagogy, Andragogy, Heutagogy, Rhizome: Who Decides What You Learn — the governance dimension this page sets alongside distribution, temporality, and epistemic role.
• Learning as a Network: Connectivism, Knowledge and Conceptual Navigation — the connectivist foundation, forms of knowing, and calculus of trust this page builds on rather than re-deriving.
• The Self-Model as a Learning Tool: Metacognition, Regulation and Agency — the self-monitoring this page's trust questions ultimately feed into.
• Inside the Rhizome: Four Social Groupings, and How Each Should Count — a closer look at the community layer this page treats as active infrastructure rather than a passive source.
• A Distributed Embedding Calculus of Trust — the relational trust model this page's ten questions extend to cover persistence and timing.
• Meaning, Ontology, and the Limits of Fixed Definitions — the computer-science sense of ontology behind a knowledge graph's schema, and what documenting one properly (scope, provenance, version, temporal validity) actually involves.

References