Why AI-native knowledge systems need to distinguish between stored knowledge and compiled knowledge.
Most AI-generated notes should probably be disposable.
This proposition may initially appear counterintuitive, particularly within the contemporary landscape of personal knowledge management (PKM), where the accumulation, preservation, and interconnection of information have long been treated as intrinsic goods. From conventional note-taking applications to increasingly sophisticated AI-assisted knowledge environments, the prevailing architectural assumption remains remarkably consistent: when a piece of information is generated, it should be saved.
Yet this assumption becomes increasingly difficult to defend when the cost of producing information approaches zero.
Consider a relatively ordinary workflow. A user imports twenty documents into a knowledge base and asks an AI assistant to summarize them. The system produces a file called summary.md. Several weeks later, another agent is instructed to synthesize the original documents together with that summary. The resulting synthesis is saved as another note. A third agent subsequently retrieves both generated artifacts, incorporates their conclusions into a broader analysis, and persists yet another document.
Every step appears individually reasonable. The agents are performing useful cognitive labor: compressing, integrating, and reorganizing information to make it more accessible.
Collectively, however, these operations introduce a structural ambiguity that conventional note-taking systems were never designed to resolve.
Which of these artifacts constitutes knowledge, and which merely represents a computation performed over existing knowledge?
The distinction is not semantic pedantry. It determines how information should be stored, updated, retrieved, trusted, and eventually discarded.
And perhaps surprisingly, database systems have been dealing with a closely related problem for decades.
The Database Analogy: Source Data and Materialized Views
In relational database architecture, a fundamental distinction exists between authoritative source records and information derived from those records.
Consider an accounting system. Individual transactions constitute the underlying ledger. A monthly revenue report, by contrast, is an aggregation computed from those transactions. Although the report may be stored for performance reasons, its values remain dependent on the records from which they were calculated.
If a transaction is corrected, the monthly report may become outdated. The appropriate response is not necessarily to edit the report manually, but to refresh the computation against the updated ledger.
This is the conceptual foundation of a materialized view.
Unlike an ordinary database view, which typically evaluates its defining query when accessed, a materialized view stores the result of a query so that it can be retrieved efficiently. Its defining property is therefore not that it is transient or physically absent from storage, but that its content is derived from identifiable inputs through a defined transformation.
The PostgreSQL documentation makes this distinction explicit: materialized views persist query results while retaining the query definition necessary for refreshing them.
The same architectural principle could be applied to AI-native knowledge systems.
A source document should be treated as a source record. A person's observation or explicitly adopted decision may constitute a canonical record of what that person observed or decided. An AI-generated summary, however, is ordinarily a transformation of existing material. A synthesis combining several summaries is another transformation, often dependent on multiple preceding transformations.
These objects may share a textual representation, but they do not share the same epistemic status.
A useful initial classification would therefore be:
- Source material: Preserved as an attributable record.
- Human observations and decisions: Preserved as canonical records of the user's statements, commitments, or judgments.
- AI-generated summaries: Treated as derived representations of source material.
- AI-generated syntheses: Treated as higher-order derived representations, with explicit dependencies.
Importantly, canonical status does not imply objective truth. A person's observation may be mistaken, and an imported document may contain inaccuracies. Canonicality establishes the authority of a record within the system's history, not the infallibility of its contents.
This distinction leads to a more general architectural proposition:
A summary is often a computation over knowledge, not knowledge itself.
Once this principle is accepted, the conventional practice of saving every generated summary as an independent, permanent note begins to look less like knowledge management and more like uncontrolled cache accumulation.
The Problem of Recursive Epistemic Amplification
The consequences extend beyond unnecessary storage consumption.
When AI-generated artifacts are repeatedly retrieved and incorporated into subsequent generations of content, the system can progressively obscure the relationship between an assertion and its original evidence.
Suppose a source document states that a proposed architectural change might reduce processing latency under certain conditions. An AI summary compresses that statement into a more definite claim: the proposed change improves latency. A later synthesis combines this summary with several related notes and concludes that the change has demonstrated performance benefits.
By the third transformation, a conditional hypothesis may have acquired the linguistic appearance of an established result.
No individual agent necessarily intended to fabricate evidence. The distortion emerged through successive operations of compression, interpretation, and recombination.
More importantly, conventional retrieval systems may subsequently encounter several documents expressing approximately the same conclusion. Without provenance awareness, those documents can appear to constitute independent corroboration, even though all of them descend from a single uncertain source.
This is a form of epistemic amplification through derivative redundancy.
The underlying issue is not simply that language models hallucinate. It is that the knowledge system permits model-generated interpretations to re-enter the evidence pool without preserving the distinction between an original assertion and its descendants.
Research on temporally evolving retrieval memory has similarly highlighted the difficulty of distinguishing superseded facts from current information through semantic similarity alone. A retrieval system may find an obsolete statement highly relevant precisely because its semantic content closely resembles the updated one.
Recursive summarization and recursive model training are not equivalent phenomena, and the former should not be casually conflated with the technical concept of model collapse. Nevertheless, both illustrate the broader importance of preserving the boundary between original evidence and machine-generated representations.
The practical danger for PKM is straightforward: a system can become increasingly confident in conclusions without acquiring any additional independent evidence.
What initially appears to be knowledge growth may instead be the multiplication of representations of the same underlying material.
From Persistent Documents to Dependency-Aware Knowledge
The more appropriate architecture may resemble a build system rather than a conventional notebook.
In software compilation, source files possess a different lifecycle from the artifacts generated from them. A compiler transforms source code into executable output; when the source changes, the corresponding build artifacts may need to be regenerated.
Developers generally do not treat compiled binaries as authoritative replacements for the source code from which they originated.
An AI-native PKM could adopt an analogous distinction between stored knowledge and compiled knowledge.
The first category contains the durable records that the system is responsible for preserving. The second contains interpretations, summaries, topic overviews, and syntheses generated from those records for particular cognitive purposes.
Under this model, an AI-generated artifact would carry a dependency specification rather than existing as an isolated document.
That specification might identify its source nodes, relevant source versions, transformation instructions, model configuration, generation timestamp, and the purpose for which the artifact was produced.
A simplified representation might look like this:
DerivedArtifact = f(SourceVersions, Instructions, ModelConfiguration)
This expression is an architectural abstraction rather than a guarantee of deterministic computation. Unlike a conventional SQL aggregation, an LLM-based transformation may produce different outputs under apparently identical conditions. Recreating the function of a summary is therefore not necessarily equivalent to reproducing its exact historical text.
That limitation matters. A system intended to support auditing or reproducibility may need to preserve generation records and selected output versions, even when the generated text itself is ordinarily disposable.
Nevertheless, dependency tracking would make several important capabilities possible.
If an underlying document changes, the system could mark dependent summaries as stale. If a source is deleted or superseded, downstream interpretations could be invalidated or flagged for review. If multiple agents request equivalent analyses, the system could reuse an existing materialization rather than generate redundant documents.
Most significantly, a retrieval engine could distinguish between independent evidence and several representations derived from the same evidence.
The resulting system would no longer operate as an undifferentiated collection of text files. It would maintain an explicit account of how its information came into existence.
This is closely aligned with the principles of data provenance formalized in the W3C PROV framework, which distinguishes entities, activities, agents, and derivation relationships.
For knowledge systems, provenance should not merely be metadata displayed after the fact. It should influence retrieval, updating, and the authority assigned to generated content.
Staleness Is a Lifecycle State, Not a User's Housekeeping Problem
Traditional PKM systems implicitly assign responsibility for informational maintenance to the user.
A note becomes outdated; the user must notice. Two documents become redundant; the user must reconcile them. A summary contradicts its original sources; the user must identify the discrepancy.
This arrangement becomes increasingly untenable when AI agents can generate derivative documents faster than their human operators can meaningfully inspect them.
The asymmetry is structural. Generation has become inexpensive, but verification, interpretation, and decision-making remain comparatively scarce human resources.
Consequently, the accumulation of AI-generated notes can produce an expanding maintenance obligation that exceeds the informational value of the notes themselves.
The appropriate response is not necessarily to build better tools for manually cleaning an ever-growing vault. It is to reconsider why so many generated artifacts were granted permanent status in the first place.
A derived artifact could instead move through explicit lifecycle states: fresh, stale, invalidated, archived, or superseded.
A fresh artifact reflects its declared inputs under its recorded generation conditions. A stale artifact remains potentially useful but no longer corresponds to the current versions of those inputs. An invalidated artifact should not be treated as current evidence. An archived artifact is intentionally retained for historical reference, while a superseded artifact has been replaced by a newer interpretation.
These distinctions enable policies that are considerably more sophisticated than indiscriminate persistence.
A frequently accessed topic overview could be refreshed when its dependencies change. An inexpensive summary could be regenerated on demand. An unused intermediate synthesis could expire automatically. A historically significant analytical report could remain permanently accessible, explicitly identified as a snapshot of an earlier state.
Such policies would transform garbage collection from an afterthought into an integral component of knowledge architecture.
The goal would not be to minimize the number of stored bytes. It would be to minimize unjustified informational persistence.
The Critical Exception: When an Interpretation Becomes a Decision
A purely computational model of AI-generated knowledge would nevertheless be insufficient.
Not every AI-assisted artifact should remain disposable indefinitely.
Suppose an AI agent evaluates several competing architectural proposals and produces a comparative analysis. The user examines its reasoning, challenges several assumptions, and ultimately selects one proposal as the basis for implementation.
At that moment, something important has changed.
The original analysis was a derived artifact. The user's adoption of a particular decision is a new historical event.
That decision cannot simply be regenerated from the original documents, because it includes an act of human judgment.
The system should therefore preserve a distinction between the analysis as a computational product and the decision as a durable commitment.
This suggests a process of canonical promotion: a derived artifact, or a specific claim extracted from it, may become part of the durable knowledge record through an explicit act of acceptance, revision, or endorsement.
Such promotion should not erase the artifact's provenance. Nor should it automatically convert every factual assertion within the analysis into verified truth. Instead, it records precisely what the user accepted, under which assumptions, and for what purpose.
This separation is particularly important for systems intended to augment human reasoning rather than silently substitute for it.
AI can generate hypotheses, identify relationships, and propose interpretations. It can also make mistakes, overlook context, or produce plausible but unjustified conclusions.
Human judgment supplies a different kind of information: a record of what was consciously evaluated, selected, rejected, or committed to action.
The distinction is therefore not merely between human-generated and machine-generated text. It is between evidence, computation, and accountable judgment.
These categories may interact, but collapsing them into a single undifferentiated notion of a note sacrifices information that an intelligent knowledge system should preserve.
Toward a Compiled Knowledge Layer
For graph-based PKM systems, this distinction has especially consequential implications.
A knowledge graph already represents information through entities and relationships. Extending that architecture to represent derivation, dependency, validity, and epistemic authority would allow the system to manage not only what information exists, but how that information should behave over time.
A source node could support several derived interpretations without those interpretations becoming independent sources of authority. A decision node could reference the evidence and analysis that informed it. A global summary could be treated as a regenerable projection of the graph rather than an ever-expanding document that must be manually maintained.
A system such as Infinite Graph could ultimately distinguish between a durable knowledge layer, a layer of explicitly adopted human interpretations, and a computational layer of AI-generated representations.
These layers need not be physically separated into different databases. Their distinction could be enforced through metadata, provenance edges, lifecycle policies, and retrieval rules.
The essential requirement is behavioral rather than cosmetic.
A source record should not disappear merely because its summary expires. A human decision should not be silently rewritten because a newer model produces a different recommendation. A derived synthesis should not acquire independent evidentiary weight simply because several agents have reproduced it.
Nor must every derivation be recomputed immediately whenever one of its inputs changes. Refresh policies can account for computational expense, relevance, urgency, and the consequences of relying on outdated information.
The architecture should also preserve exceptions. Some AI-generated artifacts are themselves historical records: a published report, a contractual deliverable, a model evaluation result, or a creative work whose exact wording matters. Their persistence may be justified independently of whether a similar artifact could be regenerated.
The principle, therefore, is not that everything written by AI should be deleted.
It is that persistence should be an explicit property of an artifact's purpose and provenance, rather than the automatic consequence of its generation.
This represents a significant departure from the storage-first assumptions inherited from conventional note-taking software.
In a world where generating text is expensive, retaining a useful synthesis is often the rational default. In a world where generating text is inexpensive and automated agents can produce thousands of overlapping interpretations, indiscriminate retention becomes a liability.
The scarce resource is no longer the ability to produce another document. It is the ability to maintain a trustworthy distinction between what was observed, what was inferred, what was decided, and what can safely be computed again.
The next generation of knowledge systems may consequently need to resemble databases, compilers, and provenance-aware dependency graphs as much as they resemble notebooks.
Your PKM may need a distinction between stored knowledge and compiled knowledge.
That distinction would change what it means to save a note, what it means to update a knowledge base, and ultimately what it means for an AI system to remember.
Because memory is not merely the persistence of information. It is also the preservation of the relationships that make information meaningful and trustworthy.
And that leaves us with a deceptively simple architectural question:
If an AI-generated note can be recreated in ten seconds, should it have the same status as something you personally observed or decided?
References and Further Reading
- PostgreSQL Documentation — Materialized Views. Technical foundations for stored query results and refresh semantics.
- W3C — PROV Model Primer. A conceptual framework for provenance, derivation, and attribution.
- Temporal Validity in Retrieval Memory: Eliminating Stale-Fact Errors for AI Agents over Evolving Knowledge (2026). A research preprint investigating temporal validity and superseded facts in retrieval-based agent memory.