Executive Summary
Most assets depreciate. Your car loses value the moment you drive it off the lot. Your laptop is worth half its purchase price within two years. Even your professional certifications expire and must be renewed.
Your knowledge graph does the opposite. Every conversation adds nodes. Every file ingestion adds structured knowledge. Every entity resolution connects previously isolated facts. The graph does not lose value over time. It gains value -- and the rate of value creation accelerates as the graph grows, because each new node has more existing nodes to connect to.
This is compound interest applied to intelligence. And like compound interest, the returns are modest in the early months and transformative over years.
But compound intelligence only works if the asset is truly yours. If the knowledge graph lives on a provider's server, the compounding accrues to the provider's retention strategy, not to your capability. If the graph lives on your device and works with any model from any provider, the compounding accrues to you -- unconditionally, permanently, and portably.
This paper examines the economics of appreciating intelligence: how knowledge graphs compound, why provider-agnostic growth creates the strongest asset, and what happens when you let compound intelligence run for years instead of months.
1. The Mathematics of Compound Intelligence
1.1 Nodes Are Not the Metric
The naive measure of a knowledge graph is node count. A graph with 10,000 nodes sounds more valuable than one with 1,000. But node count is to knowledge graphs what page count is to books -- a measure of volume, not value.
The real metric is connectivity. A knowledge graph's value is proportional to the number of meaningful connections between nodes -- the edges that link concepts, people, projects, deadlines, preferences, and insights into a navigable web of understanding.
Metcalfe's Law states that the value of a network is proportional to the square of the number of connected nodes. While the strict mathematical formulation is debated, the directional insight holds: a knowledge graph with 1,000 well-connected nodes is dramatically more valuable than one with 10,000 isolated nodes, because connectivity enables inference, suggestion, and cross-domain discovery.
1.2 The Network Effect Within a Single Brain
Knowledge graphs exhibit an internal network effect. Each new node has the potential to connect to every existing node that it is semantically related to. As the graph grows, the probability that a new piece of knowledge connects to something already in the graph increases.
In a graph with 100 nodes, a new fact about 'Project Aurora' might connect to 2-3 existing nodes. In a graph with 10,000 nodes, the same fact might connect to 15-20 existing nodes -- references to the project team, the client, the technology stack, related deadlines, previous discussions, and connected initiatives.
This means the marginal value of each new node increases as the graph grows. The 10,000th node is more valuable than the 100th node, because it has more nodes to connect to. The graph does not just grow. It accelerates.
1.3 The Compound Interest Analogy
Financial compound interest works because returns generate further returns. A $10,000 investment at 7% annual return becomes $19,672 in 10 years -- not because each year adds $700, but because each year adds 7% of the growing total.
Knowledge compounding works similarly. Each month of use does not add a fixed amount of value. It adds value proportional to the existing knowledge base, because new knowledge connects to and enriches existing knowledge. A knowledge graph that has been growing for six months absorbs new information more effectively than one that started yesterday -- because the new information has more existing context to attach to.
The implication is profound: the difference between a 6-month knowledge graph and a 24-month knowledge graph is not 4x. It is closer to 10-15x in terms of contextual utility, because the compounding has had time to work.
2. Provider-Agnostic Growth
2.1 The Growth That Transfers
A knowledge graph built over six months of daily interaction with Claude works identically when you switch to GPT or Gemini or a local Llama model. The six months of accumulated intelligence -- your domain knowledge, your preferences, your communication patterns, your project context -- transfers completely.
This is because Kent's knowledge graph does not encode any provider-specific information. Nodes are facts, preferences, entities, and relationships. Edges are connections between them. Embeddings are generated locally using all-MiniLM-L6-v2, independent of any AI provider. The schema does not reference Claude, GPT, or any other model.
When you switch providers, the knowledge graph does not change. Not a single node is modified. Not a single edge is lost. Not a single embedding is regenerated. The only thing that changes is which model receives the query -- and the query includes the same context from the same local graph regardless of which model processes it.
2.2 Contrast: The Growth That Doesn't Transfer
Now consider what happens when you switch from ChatGPT to Claude without Kent.
After six months of daily ChatGPT use, OpenAI has accumulated:
- ~200 explicit memory facts about you
- Thousands of conversation turns that have shaped the implicit behavioral model
- Custom GPTs configured for your specific workflows
- Interaction patterns that inform response style, depth, and format
When you switch to Claude, all of this stays on OpenAI's servers. You can download your conversation history as JSON, but Claude cannot import it. You can manually tell Claude your preferences, but you will forget most of them -- the implicit ones were never visible to you in the first place. You can recreate custom instructions, but the behavioral nuance built over thousands of interactions is gone.
The six months of compound intelligence? Zero transfer. You start from zero with Claude.
2.3 The Multi-Provider Compound
Kent's provider-agnostic architecture creates a unique compounding dynamic: knowledge accumulated through different providers enriches the same graph.
A user who routes complex analysis to Claude, quick lookups to Gemini Flash, and private queries to a local Llama model is building a single knowledge graph from the output of all three. The Claude-powered analysis creates high-value nodes about complex topics. The Gemini lookups create utility nodes about quick facts. The local Llama queries create private nodes that never touch a cloud server. All three contribute to the same graph, connected by the same entity resolution pipeline, searchable through the same embedding space.
No single-provider AI can offer this. By definition, a single-provider system builds knowledge only from interactions with that provider. Kent builds knowledge from every provider simultaneously -- and the compound effect applies to the aggregate, not just to any single stream.
3. What Compound Intelligence Looks Like
3.1 Month 1: The Foundation
In the first month, the knowledge graph is sparse. New facts connect to few existing nodes. The AI knows your name, your role, a few preferences, and some basic project context. Responses are generic -- similar to what any new AI assistant would produce.
The graph contains approximately 200-500 nodes, mostly from initial conversations and file ingestions. Entity resolution is beginning to identify recurring people, projects, and concepts, but the connections are thin.
Value assessment: the AI is slightly more useful than a context-free chatbot, but the difference is marginal.
3.2 Month 6: The Inflection
By month six, the graph has reached 3,000-6,000 nodes with thousands of edges. Entity resolution has identified your key clients, your major projects, your recurring collaborators, and your domain vocabulary. The AI can make cross-project connections: it notices that a technique mentioned in Project A is relevant to a problem in Project B.
Preferences have stabilized. The AI knows your preferred output format, your communication style, and your technical depth. Responses require fewer revision cycles -- the Stanford HAI finding of 0.8 cycles versus 2.3 cycles for context-naive AI becomes apparent.
Value assessment: the AI is noticeably better than a fresh one. Users at this stage report that switching providers feels unacceptable -- the context difference is too large.
3.3 Month 12: The Compound Effect
At twelve months, the graph has 10,000-15,000 nodes with dense cross-connections. The AI does not just know your current projects -- it remembers how they evolved. It can trace a client relationship from initial contact through proposal through engagement through delivery. It can suggest connections between current challenges and solutions you found months ago in a different context.
This is compound intelligence made visible. The AI is not just recalling facts -- it is synthesizing across a year of professional experience encoded in the graph. Responses are contextually rich in ways that would be impossible without the accumulated knowledge.
Value assessment: the knowledge graph has become a competitive advantage. The AI's understanding of your work exceeds what a new colleague could develop in months of shadowing.
3.4 Year 2 and Beyond: The Moat
After two years, the graph approaches 25,000-30,000 nodes with deep, multi-layered connections. The AI has institutional memory -- it remembers decisions made eighteen months ago, the reasoning behind them, and the outcomes that resulted. It can pattern-match current situations against historical ones with remarkable accuracy.
This is the point at which the knowledge graph becomes a genuine moat. No competitor can replicate it, because it encodes two years of your specific professional experience, your specific relationships, your specific domain knowledge, and your specific working patterns.
And because this graph is local and provider-agnostic, the moat belongs to you -- not to any provider.
4. The Appreciation Curve vs. the Depreciation Trap
4.1 Cloud Memory Depreciates
Cloud-held AI memory depreciates in three ways:
Policy depreciation. Providers change their memory policies. OpenAI has modified ChatGPT's memory feature multiple times since launch -- adding limits, changing retention periods, and adjusting how memories are stored. Your accumulated intelligence is subject to policy changes you have no input into.
Model depreciation. When a provider releases a new model, the implicit behavioral model built by the old model does not transfer cleanly. GPT-4o has different response characteristics than GPT-4 Turbo, which has different characteristics than GPT-3.5. The nuanced understanding built over months with one model may not apply to the next model -- even from the same provider.
Regulatory depreciation. As the Fable 5 episode demonstrated, government action can depreciate provider-held intelligence to zero overnight. Export controls, sanctions, and regulatory directives can sever your access to accumulated context with no migration path and no recovery timeline.
4.2 Local Knowledge Appreciates
Kent's local knowledge graph appreciates because it is insulated from all three depreciation vectors:
No policy risk. The graph is on your device. No provider can change its retention policy. No terms of service can limit how many facts you store. No corporate decision can modify the graph's behavior.
Model-independent. The graph's value comes from its structure -- nodes, edges, embeddings, confidence scores -- not from any specific model's behavioral patterns. When you switch from Claude to GPT, the graph's value is unchanged because the graph was never model-dependent.
Regulation-proof. A local SQLite database is not subject to export controls on AI models. Government directives target providers and their services, not personal databases on personal hardware. The regulatory risk that depreciated Fable 5 users' accumulated context to zero does not apply to locally stored knowledge.
5. Building the Appreciating Brain
5.1 Ingestion as Investment
Every file you drop into Kent, every conversation you have, every connector query you run is an investment in a compound asset. The key is consistency: the compound effect requires sustained input over time.
A professional who ingests 5-10 documents per week, has 3-5 substantive AI conversations per day, and connects 2-3 external data sources is building a knowledge graph at approximately 200-400 new nodes per week. Over a year, this produces 10,000-20,000 nodes -- a substantial knowledge base that encodes a year of professional activity.
The investment cost is negligible -- the user is already doing the work (reading documents, having conversations, querying databases). Kent captures the knowledge as a side effect of normal workflow. The only 'cost' is having Kent running, which requires no behavioral change.
5.2 The Workspace Strategy
Kent's workspace model allows users to segment their knowledge graph by domain -- a workspace for each client, each project, or each area of expertise. This segmentation enables focused retrieval (the AI searches within the relevant workspace rather than the entire graph) while maintaining the option of cross-workspace search when broader context is needed.
The workspace strategy also enables selective sharing. A consultant can maintain separate client workspaces that never cross-contaminate, while their 'Professional Development' workspace accumulates domain expertise from all client work.
5.3 The Carry-Forward Principle
When a project ends, the knowledge graph does not delete itself. The project workspace is archived but remains searchable. Six months later, when a similar project begins, the AI can draw on the archived workspace's knowledge -- methodologies, solutions, client preferences, lessons learned -- without the user needing to remember or manually retrieve any of it.
This carry-forward principle is what transforms a knowledge graph from a tool into a career asset. A professional who uses Kent for five years has a searchable, connected record of their entire professional output -- every project, every document, every insight, every decision. That record is portable, private, and permanent.
Conclusion
Every asset either appreciates or depreciates. There is no stasis.
Cloud-held AI memory depreciates -- eroded by policy changes, model updates, and the ever-present risk of regulatory disruption. You invest your time and attention into building a provider's understanding of you, and that investment can be wiped out by a terms-of-service update or a government directive.
Local, provider-agnostic knowledge appreciates -- growing more valuable with every interaction, transferable across providers, immune to policy changes, and compounding over years. The investment accrues to you. The returns are permanent.
The difference between these two outcomes is architecture. Same user. Same work. Same daily interactions. One architecture makes you more dependent with every passing month. The other makes you more capable.
Choose the one that compounds in your favor.
References
- Stanford HAI. (2025). 'Context-Aware AI and Output Quality in Enterprise Settings.'
- Metcalfe, R. (2013). 'Metcalfe's Law After 40 Years of Ethernet.' *IEEE Computer*, 46(12), 26-31.
- Kent. (2026). 'Internal Usage Analytics: Knowledge Graph Growth Rates and Connectivity Metrics.'
- Bain & Company. (2025). 'Consumer Willingness to Pay for AI Personalization.'
- McKinsey & Company. (2025). 'The State of AI in 2025: Global Survey Results.'
Published by Kent Research, July 2026.