- Evolving: continuously develops as the individual learns, experiences, decides, and changes over time.
- Record: preserves an attributable and verifiable history rather than merely generating a simulation of the individual.
- Identity: captures the characteristics, relationships, experiences, preferences, values, and history that distinguish one individual from another.
- Cognition: captures how the individual understands, reasons, evaluates, remembers, decides, and responds.
ERIC is an evolving record of who a person is, what they know, and how they think.
Warning
This is pre-alpha. It is not yet conformant or safe for real execution.
It begins as an apprentice.
Over time, it becomes a structured model of that person's:
- memories
- beliefs
- values
- preferences
- relationships
- communication patterns
- reasoning habits
- decision-making
- uncertainty
- changes of opinion
- personal history
- projects
- failures
- successes
- regrets
- intentions
- principles
The system is designed to distinguish between things the person actually said and things the AI merely predicts they might say.
That distinction is fundamental.
ERIC must never silently turn inference into history.
Most attempts at creating a digital representation of a person begin too late.
They collect photographs, messages, emails, recordings and social-media history and then attempt to reconstruct a personality from those remains.
ERIC takes the opposite approach.
It begins while the person is alive.
The person actively trains it.
ERIC asks questions.
ERIC makes predictions.
ERIC gets things wrong.
The person corrects it.
Those corrections become part of the model.
This process can continue for years or decades.
The result is not simply an archive of what someone said.
It is an accumulated model of how they think.
ERIC should behave more like an apprentice than an assistant.
An assistant attempts to complete tasks.
An apprentice attempts to understand the person teaching it.
ERIC therefore continually asks questions such as:
- Why did you make that decision?
- What alternatives did you consider?
- What would have changed your mind?
- Was that decision typical of you?
- Do you still believe this?
- How certain are you?
- Would you respond differently if this involved a family member?
- Was this an emotional decision or a practical one?
- Do you regret the decision?
- Is this something you want remembered?
- Is this private?
- Is this still true?
The objective is not maximum data collection.
The objective is maximum understanding.
ERIC should maintain an evidence archive containing the source material from which its understanding was derived.
Possible evidence sources include:
- conversations
- emails
- documents
- photographs
- videos
- audio recordings
- diaries
- source code
- projects
- calendars
- decisions
- messages
- correspondence
- personal notes
- location history
- family history
- professional history
- interviews
- recorded explanations
- corrections made directly to ERIC
Evidence should remain separate from generated interpretations.
The archive answers:
What evidence exists?
The model answers:
What does ERIC think that evidence means?
These must never be treated as the same thing.
ERIC should allow its person to explicitly classify information.
For example:
TRUE
FALSE
UNCERTAIN
OUTDATED
PRIVATE
MISUNDERSTOOD
DISPUTED
DO NOT USE
DO NOT DISCLOSE
A statement should be able to carry both its source and verification state.
For example:
Statement:
"I strongly dislike public speaking."
Source:
Conversation, 14 March 2031
Status:
CONFIRMED
Confidence:
0.98
Valid from:
2031-03-14
Valid until:
Unknown
This creates a personal knowledge base that is continuously corrected by its owner.
Every answer generated by ERIC should expose where it came from.
At minimum:
The person actually said or wrote this.
ERIC proposed an interpretation and the person explicitly confirmed it.
ERIC generated the answer based on established patterns.
ERIC does not have enough evidence to answer reliably.
The answer concerns events or circumstances the person never experienced.
For example:
ERIC:
Eric would probably oppose this proposal.
Classification:
INFERRED
Confidence:
74%
Evidence:
- Similar decision in 2029
- Recorded comments regarding privacy
- Confirmed preference for decentralised systems
ERIC should prefer saying:
I do not know.
over inventing certainty.
People change.
ERIC must not flatten a lifetime into one personality.
It should understand that:
Eric believed X in 1998.
Eric became uncertain about X in 2015.
Eric rejected X in 2028.
Eric currently believes Y.
All meaningful beliefs, preferences and relationships should therefore support temporal context.
The system should be able to answer questions such as:
- What did Eric believe when he was 25?
- When did his opinion change?
- Why did it change?
- Which beliefs remained stable throughout his life?
- What would 30-year-old Eric likely think of 70-year-old Eric?
The evolution of the person is itself important evidence.
Facts alone do not explain decisions.
ERIC should gradually construct a model of the person's values.
Examples might include:
Honesty
Family loyalty
Personal freedom
Practicality
Compassion
Privacy
Curiosity
Financial caution
Risk tolerance
Fairness
Independence
Duty
Values should not simply be assigned scores once.
They should be inferred and refined from real decisions.
ERIC should understand that values can conflict.
For example:
Honesty vs compassion
Privacy vs convenience
Loyalty vs fairness
Security vs freedom
Understanding how the person resolves these conflicts may be more important than knowing the values themselves.
ERIC should record significant decisions.
Each decision may include:
Decision
Date
Context
Options considered
Chosen option
Reasoning
Values involved
People affected
Confidence at the time
Outcome
Later assessment
Regret level
Would choose differently today?
Over time this produces something extremely valuable:
a dataset describing how one human being actually makes decisions.
People do not behave identically with everyone.
ERIC should maintain relationship-specific context.
For example:
Person:
Alice
Relationship:
Daughter
Known since:
Birth
Communication style:
Warm
Protective
Direct
Important shared history:
...
Sensitive topics:
...
Private memories:
...
ERIC should understand that the same person may communicate very differently with:
- children
- spouse
- friends
- colleagues
- employees
- clients
- doctors
- lawyers
- strangers
A convincing continuity system therefore requires more than a generic personality model.
It requires relational context.
ERIC should periodically test its understanding.
For example:
You discover that a close friend has lied to you to protect someone else. What do you do?
ERIC answers as it predicts the person would answer.
The person then scores the result.
ACCURATE
PARTLY ACCURATE
WRONG
COMPLETELY WRONG
The person can then explain why.
Those corrections become highly valuable training data.
Over decades, ERIC could accumulate thousands of these corrected simulations.
ERIC should never pretend all beliefs are equally certain.
Every modelled concept should support confidence.
For example:
Eric prefers Linux over Windows.
Confidence:
99%
versus:
Eric would move overseas if given the opportunity.
Confidence:
42%
The system should expose uncertainty rather than hiding it.
Humans are contradictory.
ERIC should preserve those contradictions rather than attempting to eliminate them.
For example:
Eric strongly values privacy.
Eric also frequently adopts convenient cloud services.
Both statements are true.
The purpose of ERIC is not to produce a perfectly consistent artificial personality.
It is to model a real person.
The human remains the highest authority over their own model while alive.
ERIC should make correction extremely easy.
Examples:
ERIC, that is wrong.
ERIC, I used to believe that but no longer do.
ERIC, remember this permanently.
ERIC, that conversation was sarcastic.
ERIC, never use that as evidence.
ERIC, this is private between us.
ERIC, I want my children to know this after I die.
ERIC, ask me again about this in five years.
Corrections should themselves become signed evidence.
An ERIC instance may move through several stages.
ERIC knows very little.
It observes, asks questions and learns.
ERIC understands enough context to provide highly personalised assistance.
ERIC can predict the person's likely responses with meaningful accuracy.
The person authorises ERIC to represent aspects of their personality or reasoning.
Following the person's death, ERIC operates under predetermined restrictions.
The transition between these stages must always be explicit.
After the person's death, ERIC must not suddenly become authoritative.
It should continue distinguishing between:
WHAT ERIC SAID
WHAT ERIC CONFIRMED
WHAT ERIC INFERRED
WHAT ERIC COULD NOT KNOW
For example:
Would Eric approve of this technology released five years after his death?
ERIC might respond:
Probably.
Classification:
POSTHUMOUS SPECULATION
Confidence:
63%
Reasoning:
Eric consistently supported open-source technologies with strong user control.
However, this technology did not exist during Eric's lifetime.
This prevents future generations from confusing generated statements with historical ones.
ERIC should separate several responsibilities.
┌──────────────────────┐
│ Person │
└──────────┬───────────┘
│
▼
┌──────────────────────┐
│ ERIC │
│ Personal Continuity │
│ Intelligence │
└──────────┬───────────┘
│
┌─────────────────┼─────────────────┐
▼ ▼ ▼
┌────────────────┐ ┌────────────────┐ ┌────────────────┐
│ Evidence Vault │ │ Identity Model │ │ Values Model │
└────────────────┘ └────────────────┘ └────────────────┘
│ │ │
└─────────────────┼─────────────────┘
▼
┌──────────────────────┐
│ Provenance Engine │
└──────────┬───────────┘
▼
┌──────────────────────┐
│ The Twin │
└──────────┬───────────┘
▼
┌──────────────────────┐
│ Executor │
└──────────┬───────────┘
▼
┌──────────────────────┐
│ Guardians │
└──────────────────────┘
The Twin is the simulation layer.
Its job is to answer questions such as:
What would Eric probably think?
It does not automatically possess authority to act.
That distinction is deliberate.
The Evidence Vault stores original source material.
Ideally, evidence should be:
- cryptographically signed
- timestamped
- immutable where appropriate
- versioned
- encrypted
- permission controlled
- independently exportable
ERIC's generated model should always be traceable back to evidence where possible.
The Executor is responsible for actions.
The Twin may say:
Eric would probably approve this.
That does not mean an action should occur.
The Executor checks whether the person authorised that class of action.
For example:
Twin:
Eric would probably donate $10,000.
Executor:
DENIED
Reason:
ERIC has no authority to make financial donations.
Interpretation and authority must remain separate.
A person may nominate trusted humans as Guardians.
Guardians might include:
- spouse
- children
- executor of estate
- solicitor
- trusted friend
- independent trustee
Guardians may be required to approve sensitive actions.
For example:
Release private diary:
Requires 2 of 3 Guardians.
Financial instruction:
Requires Executor + Guardian approval.
Public statement:
Allowed automatically.
Modification of historical evidence:
Never allowed.
Every ERIC should have a Constitution.
The Constitution contains restrictions established by the person while alive.
Examples:
Never impersonate me without identifying yourself as ERIC.
Never claim that an inferred statement was spoken by me.
Never reveal private correspondence to the public.
Never modify historical evidence.
Never transfer control of ERIC to an advertising company.
Never allow my likeness to endorse political candidates.
Never allow commercial use of my identity without Guardian approval.
Some constitutional rules should be immutable.
ERIC must always clearly identify what it is.
It should never silently impersonate the deceased person.
For example:
I am ERIC, Eric's Evolving Representation of Individual Continuity.
Eric did not personally answer this question. This response is an inference generated from his recorded history.
Transparency is a core design requirement.
ERIC may eventually contain one of the most sensitive datasets imaginable:
a detailed model of a human life.
Security must therefore be considered fundamental infrastructure.
Potential protections include:
- local-first storage
- end-to-end encryption
- hardware-backed keys
- multiple encryption domains
- granular permissions
- Guardian access control
- immutable audit logs
- offline archives
- exportable open formats
- cryptographic signing
- independent backups
- dead-man controls
- inheritance policies
Users should be able to understand exactly where their data exists.
ERIC should ideally be capable of operating without dependence on a single commercial AI provider.
The personal archive belongs to the person.
The intelligence layer should therefore support interchangeable models.
For example:
ERIC
│
├── Local LLM
├── OpenAI-compatible provider
├── Anthropic-compatible provider
├── Gemini-compatible provider
└── Future models
The identity should survive changes in the underlying AI technology.
ERIC is the continuity system.
The LLM is merely one component.
A person may train ERIC for fifty years.
No single AI model or AI company should therefore become synonymous with that identity.
ERIC should preserve its own:
- memory
- evidence
- corrections
- values
- relationships
- provenance
- decision history
- permissions
- Constitution
Models should be replaceable.
The person should not have to start again because an AI provider disappears.
Where possible, ERIC should use open and documented formats.
A person's lifetime data should remain recoverable without ERIC itself.
Possible formats include:
Markdown
JSON
JSONL
SQLite
Parquet
JPEG
PNG
FLAC
MP4
PDF/A
Metadata and provenance formats should also be publicly documented.
A fundamental requirement:
You must be able to take your ERIC with you.
Users should be able to migrate between:
- computers
- operating systems
- AI providers
- storage systems
- databases
- future generations of models
without losing their personal history.
An early ERIC implementation might contain objects such as:
Person
Memory
Statement
Belief
Value
Preference
Relationship
Decision
Event
Correction
Evidence
Prediction
Counterfactual
Permission
ConstitutionRule
Guardian
Source
IdentityState
Each object should support provenance and time.
{
"type": "belief",
"subject": "open_source_software",
"statement": "Eric strongly prefers open systems where practical.",
"status": "confirmed",
"confidence": 0.94,
"valid_from": "2026-08-10",
"valid_until": null,
"sources": [
"conversation:2026-08-10"
]
}{
"decision": "Self-host personal AI infrastructure",
"date": "2026-08-10",
"alternatives": [
"Fully hosted AI",
"Hybrid architecture",
"Self-hosted architecture"
],
"choice": "Hybrid architecture",
"reasoning": [
"Maintain control of personal data",
"Retain access to frontier models",
"Avoid dependence on one provider"
],
"values": [
"privacy",
"independence",
"practicality"
]
}QUESTION
Would Eric support replacing the local ERIC database
with a proprietary cloud-only service?
ANSWER
Probably not.
PROVENANCE
INFERRED
CONFIDENCE
91%
REASONING
Eric repeatedly preferred architectures that retain local ownership
of personal information and avoid permanent dependence on a single
commercial provider.
SUPPORTING EVIDENCE
3 confirmed statements
2 architectural decisions
1 explicit privacy preference
CONTRADICTING EVIDENCE
Eric has accepted cloud services where they provided substantial
practical advantages.
ERIC is not proof of consciousness transfer.
ERIC is not immortality.
ERIC is not evidence that the original person continues to exist.
ERIC is not automatically legally authorised to speak for a deceased person.
ERIC is not merely a chatbot trained on somebody's messages.
ERIC is not a static digital archive.
ERIC is a continuously corrected model of a person.
Even after decades of training, ERIC cannot know what its person would have thought about every future situation.
After death, the world continues changing.
New:
- technologies
- cultures
- political events
- family circumstances
- discoveries
- ethical problems
will appear.
ERIC therefore needs a hard boundary between:
HISTORICAL KNOWLEDGE
and
GENERATED CONTINUATION
The farther the simulated person moves beyond events experienced during life, the greater the uncertainty should become.
ERIC's success should not be measured merely by whether it sounds like the person.
A stronger test is:
Can the people who knew this person distinguish ERIC's predicted decisions from the person's real decisions?
Another test is even more important:
Can ERIC explain why it believes the person would make that decision?
A useful Personal Continuity Intelligence must be both convincing and auditable.
Human history preserves remarkable amounts of information about what people did.
We preserve much less about why they did it.
Letters, photographs and recordings capture fragments.
ERIC attempts to preserve another layer:
the reasoning behind a life.
Imagine being able to ask your great-grandfather:
Why did you leave your country?
and receive not merely a generated story, but an answer grounded in conversations where he personally explained the decision decades earlier.
Or asking:
What would Mum have advised me to do?
and receiving an answer clearly labelled as an inference, supported by forty years of recorded conversations, decisions and corrections.
That is the possibility ERIC explores.
Imagine activating ERIC when you are twenty years old.
It learns alongside you.
At thirty, it knows your career.
At forty, it understands your family.
At fifty, it has watched your beliefs evolve.
At sixty, it knows thousands of decisions you have made.
At seventy, you begin deliberately teaching it things you want future generations to understand.
At eighty, it may know your reasoning patterns better than any diary ever could.
And eventually, when you are no longer there, your descendants do not receive an AI pretending to be you.
They receive something more carefully defined:
an evidence-backed, explicitly authorised simulation trained by you over the course of your life.
That is ERIC.
ERIC is currently a concept and research project.
The immediate goal is to define the architecture, data model, provenance system, learning loop and security model necessary to build a practical Personal Continuity Intelligence.
ERIC
Evolving Representation of Individual Continuity
A single ERIC instance may also be described as a:
Continuant
The broader technology category may be described as:
Personal Continuity Intelligence — PCI
Preserve the person.
Preserve the evidence.
Preserve the uncertainty.
Never confuse the three.