My central thesis for the next five to ten years is that we will not train a complete model for every person or company. Instead, individuals and organizations will attach a private cognitive layer to shared base models.

The base model supplies general intelligence. Personal data supplies identity and continuity. Enterprise data supplies domain knowledge and operating rules. Agent systems turn those capabilities into action.

1. From one model for everyone to layered intelligence

Today’s general models use one shared parameter set for nearly everyone. They can answer questions about public knowledge, but they do not naturally understand a particular person or organization:

  • a user’s long-term goals, values, and communication style;
  • years of work history and the reasons behind past decisions;
  • private files, messages, photos, and behavioral data;
  • a company’s internal language, permission structure, and tacit knowledge;
  • a stable identity that can be corrected across many interactions.

A future AI system is more likely to be assembled from several layers:

1
2
3
4
5
6
7
8

AI capability
  = Base Model
  + Personal Layer
  + Domain / Company Layer
  + Real-time Context
  + Memory
  + Tools
  + Identity & Permission

The base model provides language, reasoning, coding, and broad world knowledge. The personal layer represents preferences, history, relationships, style, and goals. The domain layer adds industry methods, company knowledge, and operating rules. Context describes the current task; memory maintains continuity; tools act on external systems; identity and permission define whom the agent represents and what it may do.

This is not merely a larger prompt. It is a persistent private intelligence system built around a replaceable general model.

2. A critical asymmetry: intelligence is advancing faster than memory

Current models already match or exceed ordinary human performance on many measurable tasks in knowledge, reasoning, programming, and content generation. They can read large amounts of material, connect information across domains, and produce strong analyses within seconds. If intelligence is measured as “Can the system solve this task in one attempt?”, the frontier is rapidly approaching—and sometimes crossing—human performance.

But this task intelligence is not the whole of human intelligence. One of the clearest remaining weaknesses is reliable, continuous, long-term memory.

A person does not need to reread ten years of conversations at every meeting to remember who they are, how they relate to others, and why important decisions were made. Human memory is imperfect, but it is naturally connected to identity, time, context, and causality. Today’s language-model systems are different:

  • the model does not naturally retain an experience after a conversation ends;
  • even a very large context window is temporary working space for one inference process;
  • retrieval systems can miss relevant information or surface facts that are no longer valid;
  • models struggle to distinguish consistently between explicit user statements, system inferences, and their own generated claims;
  • false memories often lack dependable provenance, confidence, correction, and forgetting mechanisms;
  • accumulated memory is rarely portable when the user changes models or platforms.

The result is an asymmetric system:

1
2

Single-task intelligence: rapidly approaching or exceeding humans
Continuous long-term memory: still substantially behind humans

This is why a stronger base model does not automatically become a genuine personal assistant. A system may solve difficult mathematics yet remain a tool that has to “meet you again” whenever it cannot recover what you have been doing, why you did it, and which facts have changed.

The next major step for personal AI is therefore not intelligence alone. It is giving already capable reasoning systems a durable memory architecture that decides what deserves to be remembered, understands temporal validity, preserves sources and relationships, corrects errors, and lets the user move or delete memory.

The rest of this essay follows from that asymmetry:

1
2
3
4
5
6
7
8
9

Single-task intelligence becomes abundant and replaceable
        ↓
Long-term memory becomes the scarce capability
        ↓
Memory must connect to private personal and organizational data
        ↓
The private layer requires identity, permissions, tools, and evaluation
        ↓
Competition shifts from models to persistent cognitive systems

3. Inference one: abundant intelligence turns base models into infrastructure

If many models become good enough at a broad range of single tasks, access to reasoning itself becomes purchasable and replaceable infrastructure. Base models may come to resemble operating-system kernels, database engines, or cloud platforms. A small number of organizations will train frontier models, open-weight communities will offer deployable alternatives, and users will choose a substrate according to capability, cost, latency, privacy, and compliance.

“Open-weight” is often more precise than “open source,” because model licenses and training transparency vary. Regardless of terminology, cheaper inference, stronger local hardware, and efficient adaptation will make private deployment increasingly practical.

The base model remains important, but once single-task intelligence is no longer the only bottleneck, it is no longer the complete product. Long-term value shifts toward accurate and portable memory, user-controlled data, reliable tools, revocable permissions, and continuous adaptation from feedback. Intelligence can come from the base model; continuity cannot come from the base model alone.

4. Inference two: a personal LoRA cannot close the memory gap by itself

If memory is the central gap, training a personal LoRA and attaching it to a base model is an intuitive response. It solves part of personalization, but it cannot carry an entire personal memory.

LoRA freezes pretrained weights and learns small low-rank matrices for adaptation. It is well suited to stable patterns such as writing style, response preferences, domain terminology, and recurring decision frameworks. It is poorly suited to storing every changing personal fact.

When a user moves home, a company changes a price, or a project makes a new decision, retraining parameters is slow and makes provenance and deletion difficult. Non-parametric memory such as RAG is better for mutable, attributable facts. Systems inspired by MemGPT show how information can move between external storage and a limited context window.

Different information belongs in different substrates:

Information Better substrate
Stable style and behavioral tendencies LoRA / adapter / preference model
Changing personal or company facts RAG / database / file system
Entities, relationships, and time Knowledge graph / event store
Current task state Context window / working memory
Long-term habits and feedback Memory system / user model
Business operations Tools / APIs / workflows
Authorization and compliance Identity / policy engine

A more accurate formula is therefore:

1
2
3
4
5
6
7
8

Personal AI
  = Personal Adapter
  + Long-term Memory
  + Real-time Context
  + User Model
  + Tools
  + Permission System
  + Continuous Evaluation

LoRA may implement one “personal matrix,” but a personal assistant is not a matrix file. Parameters learn stable patterns, external memory stores changing experience, and permissions constrain how that memory may be used. Together they form a cognitive system that continuously serves its owner.

5. Inference three: personal data becomes a portable cognitive layer

Because public model parameters cannot create personal continuity by themselves, personal history must persist outside the model in a form the model can use. Each person may accumulate an AI-readable but user-controlled collection of digital assets: documents, email, photos, conversations, notes, search and reading history, work products, decisions, health data, relationships, goals, risk preferences, and records of accepting or correcting AI output.

A useful assistant must know more than facts about its user. It must understand each fact’s source, time range, confidence, and permitted use. It should answer:

  • Where did this memory come from?
  • Is it still valid?
  • Which application may use it?
  • Has the user corrected or revoked it?
  • May it inform an answer, or also authorize an action?

The personal data layer is therefore not merely a passive knowledge base; it is a systematic compensation for the model’s memory limitations. It needs provenance, versioning, expiry, correction, and real deletion—not an ever-growing, opaque user profile.

6. Inference four: enterprises need organizational memory, not only knowledge bases

Organizations have the same memory gap. Their scarce knowledge often lies in why a decision was made, which earlier attempts failed, who has authority, and how a rule changed over time. Traditional knowledge bases store documents but rarely preserve a complete decision history. An enterprise AI layer must become organizational memory. The general model performs language and reasoning, while private layers supply knowledge bases, product and customer data, operating procedures, terminology, project history, organizational roles, risk policies, and tools connected to internal systems.

Rather than one enormous company adapter, an enterprise may maintain modules loaded according to task and identity:

1
2
3
4
5
6
7
8

General Base Model
├── Industry Adapter
├── Company Knowledge Layer
├── Legal & Compliance Policy
├── Customer Service Adapter
├── Software Engineering Tools
├── Team Memory
└── Personal Layer

At runtime, an employee’s system might combine:

1
2
3
4
5
6

Base Model
  + Industry Capability
  + Company Capability
  + Role Capability
  + Team Context
  + Personal Capability

This is a stricter interpretation of “each company trains its own matrix.” Organizations will compress knowledge, behavior, and operational ability into loadable, composable, replaceable, and evaluated private modules. Only some of those modules will be parametric; many will live in retrieval, memory, policy, and tool systems.

7. Inference five: memory creates identity, and identity enables agent collaboration

Once a personal assistant has continuous personal memory and a company assistant has continuous organizational memory, they become more than temporary chat sessions: they can act as persistent representatives of a person and an organization. Software interaction may then move from humans opening every application toward personal agents collaborating with enterprise agents and tools.

For travel planning, a personal assistant knows budget, diet, schedule, and risk preferences. Airline and hotel agents expose current inventory and policies. The personal agent searches, compares, and assembles options within its mandate, while payment or consequential commitments still require the user’s confirmation.

Inside a company, an employee assistant understands an individual’s role, a team assistant maintains project state, and a company assistant understands policy and resources. Multiple constrained agents collaborate instead of one omnipotent assistant receiving all data and permissions.

Future protocols will therefore need to carry more than natural language. They must convey identity, authorization, provenance, available actions, and responsibility boundaries. Without verifiable identity and memory provenance, agent-to-agent interaction cannot establish durable trust.

8. The resulting bottlenecks: data, permissions, and portability

Following this argument, the bottleneck shifts from “Can the model answer?” to “Who owns the memory, can it be trusted, and how may it be used?” Model intelligence still matters, but this future depends on solving a deeper set of infrastructure problems.

Ownership and portability

Can a user move memory, preferences, and tool configurations when changing base models? If the private layer cannot leave one platform, personal AI becomes a powerful form of vendor lock-in.

Privacy and minimum exposure

Centralizing a person’s complete life history creates exceptional risk. Sensitive workloads will need local devices, home servers, or enterprise environments alongside encryption, data classification, trusted execution, federated learning, and differential privacy. These protections are not free: NIST’s work on privacy-preserving federated learning emphasizes the performance and engineering trade-offs of stronger protection.

Reading, reasoning, and execution must be separate

1

Read Permission ≠ Reasoning Permission ≠ Execution Permission

Access to a financial goal does not authorize sending it to a third party or transferring money. Authorization must specify scope, target, duration, revocability, and audit records.

Memory must support correction and forgetting

An incorrect long-term memory can distort future decisions repeatedly. Systems need source visibility, confidence, user correction, expiry, and verifiable deletion.

Adapter composition is not automatically safe

Stacking personal, company, industry, and role adapters can create conflicts, regressions, or unpredictable behavior. Routing, precedence, conflict detection, safe merging, cross-model portability, and regression evaluation remain important technical problems.

9. An infrastructure market built around the memory gap

If single-task intelligence becomes standardized while memory, identity, and permissions remain scarce, value will spread beyond the model itself:

  1. Personal Data Vaults for storing and authorizing digital assets;
  2. AI Memory Operating Systems for writing, retrieving, correcting, forgetting, and moving memory;
  3. Adapter registries and marketplaces for domain capability modules;
  4. Model routers that select a model by capability, price, latency, and privacy;
  5. Agent identity and permission systems for mandates, expiry, and audits;
  6. Personal evaluation systems that measure fit for a particular user;
  7. Private AI compute on devices, home servers, and enterprise clouds;
  8. Interoperability layers for moving memory, adapters, tools, and identity across models.

The industry currently competes over who has the strongest model. A more consequential future competition may be over who manages a user’s long-term memory, permissions, and digital identity.

10. A possible five-to-ten-year path

The main trajectory is not that models suddenly gain perfect memory; external systems progressively supply continuity. This transition is likely to occur in stages.

Stage one: context temporarily compensates for memory. Personal and enterprise systems primarily use long context, RAG, tool connections, and simple memory. Data remains fragmented across applications.

Stage two: temporary memory becomes a persistent private layer. Assistants gain cross-application memory, preference models, company knowledge, and explicit permissions. Open-weight models and local inference handle more sensitive work.

Stage three: private memory becomes modular and portable. Capabilities become less dependent on one model. Runtimes choose a substrate, load modules, and route tasks dynamically.

Stage four: persistent agent-to-agent collaboration. Personal, enterprise, and public-service agents with continuous identities interact through standard protocols. Users manage goals, boundaries, and confirmation points rather than every interface.

This is a directional thesis, not a guaranteed timetable. Closed ecosystems, fragmented formats, security failures, and regulation may slow progress substantially.

Conclusion: intelligence from the model, identity from the private layer

The argument does not begin with “train one model for every person.” It begins with the observation that as single-task intelligence grows stronger, long-term memory becomes the critical constraint on personal and enterprise assistants. The resulting thesis is:

Future models will become general computing substrates onto which people and organizations attach their own data and capabilities. Their differences will be represented not by retraining complete models, but by portable adapters, memory, knowledge stores, permissions, and tools.

Or more simply:

The base model provides intelligence, private data provides identity, and the agent system provides agency.

Models may become increasingly replaceable. A private intelligence layer accumulated over years—one that understands its owner, connects to tools, and operates within explicit authority—will be much harder to replace. The strongest moat may not be model parameters, but a persistent digital self that users can own and carry.

That is both the opportunity and the danger. A user-owned private layer can amplify individual capability. A platform-owned layer can become an unprecedented concentration of data and power.

References