OpenKedge
KedgeFlow
PCI research programLong-lived AI agents
Persistent Cognitive Identity

Governed identity for AI agents.

Persistent Cognitive Identity (PCI) studies how a long-lived AI agent keeps an accountable history as its model, tools, or runtime change. Its proposed continuity architecture connects accepted state, provenance, consent, and authority across versions.

A model can sound familiar while carrying stale memories or permissions it never earned. PCI asks which records a successor may rely on, which claims it can make, and which grants must be reviewed before it acts again.

The identity-to-authority boundary

A successor model does not inherit an agent's authority merely by sounding like it.

PDDS governs proposed actions at execution; PCI governs the identity state and permissions that an agent may carry into its next version.

See How Continuity WorksRead the ResearchExplore PDDS
Agent Identity & Continuity
Operator-Controlled Reference Architecture
Architecture intent·PCI Continuity Kernel
Evidence:Bounded model
Disposition:Commit / review

A Continuity Kernel is proposed to activate accepted changes against an exact predecessor while recording state, authority, lineage, and disposition together.

Accepted branch head and predecessor checks
Typed provenance and appropriate uncertainty
Operator-controlled custody and disclosure
Authority changes require explicit review
Accepted
Branch head
Scoped
Authority
Typed
Provenance
Review
Succession
At a glance

What is PCI?

Persistent Cognitive Identity is a research architecture for governing the state and authority of a long-lived AI agent. It records which experiences, claims, permissions, and relationships were accepted, then evaluates what may continue after a model or runtime change.

Acronym
PCI
Full term
Persistent Cognitive Identity
Applies to
Long-lived AI agents and changing model runtimes
Core question
What state and authority may a successor carry forward?
From one version to the next

How PCI applies to persistent AI agents

An agent may work across months, tools, and model releases. Continuity needs more than replaying a chat log: accepted state must be traceable, a proposed successor must be assessed, and authority must remain under human or institutional control.

  1. 01

    Record accepted state

    Track an agent's accepted experiences, claims, commitments, and authority with provenance and explicit scope.

    Continuity Kernel
  2. 02

    Evaluate a change

    When a model or runtime changes, check lineage, supported claims, consent, and the limits of the proposed successor.

    Identity tests
  3. 03

    Reauthorize the successor

    An institution decides which state and permissions transfer, which need review, and which should remain withheld.

    Governed succession

Reference flow. A transition can be rejected, quarantined, deferred, or accepted with narrower authority.

PCI Architecture Stack · Governed Agent Continuity

Reference architecture • Conditional results • Operator-defined custody

L7Sovereign Digital Talent & Workforce Layer
L6Governance & Institutional Custody Plane
L5Embodiment, Relationships & Authority Scope
L4Constitution & Epistemic Boundedness Engine
L3–L2Memory Synthesis & Evidentiary Belief Graph
L1Transactional Continuity Kernel (Beyond Memory)
BASESwappable AI Substrates (Decoupled Reasoning Engines)
PCI identity evaluation questions
HRFT · SIT · CCT
HRFT TestHuman-Referential

Evaluates fidelity to a declared human source when an agent represents that person.

SIT TestSituated Identity

Tests whether behavior can be attributed to a specific recorded lineage rather than a plausible persona.

CCT TestCognitive Continuity

Evaluates whether a proposed state transition preserves authorized lineage and bounded authority.

The problem

Most AI systems preserve only fragments of identity.

A name, prompt, memory store, checkpoint, or key can help identify an agent. None alone establishes which past state is accepted, which claims are supported, or which permissions remain valid.

When a model is upgraded, which memories are still authentic? When an agent moves to a new provider, who decides what authority transfers? When two copies of the same system run in parallel and then try to reconcile, which one is the original?

These are practical engineering questions for long-lived assistants, enterprise agents, digital twins, and embodied systems that may outlast their first model or hardware platform.

What PCI proposes

Identity as a governed lineage across change.

PCI models operational identity as a governed lineage of accepted experiences, memories, beliefs, relationships, and permissions. The architecture proposes rules for evaluating each change rather than treating a copied snapshot as proof of continuity.

Experience

A tamper-evident record of captured events and observations, within a declared collection boundary.

Memory

Episodic, semantic, procedural, and affective state derived from recorded experience.

Beliefs

Revisable claims with dependencies and evidence status, including uncertainty and conflict.

Constitution

The identity's core values, self-representation, and versioned development policy.

Relationships

Roles, obligations, fiduciary duties, and commitments to specific people and organizations.

Embodiment

Physical capabilities, sensor access, environmental constraints, and hardware affordances.

Governance

Custody rules, disclosure controls, consent state, active authority grants, and succession policy.

Why it matters

Continuity requires an authorized lineage.

Similar behavior does not establish an authorized identity transition. A copy may hold the same memories yet lack the same consent or authority. A system may sound like a person while lacking evidence for that person's lived experience—and should be able to say so.

PCI proposes ways to distinguish state that must remain stable from state that may evolve. The current program includes specified provenance and transition controls; consent, succession, custody transfer, and semantic merge remain broader architectural and research work.

For model migration, branching, merging, or embodiment transfer, PCI asks: was the change authorized, is the lineage preserved, and are the authority boundaries correct?

Scope and limits

A defined operational scope.

PCI does not claim to prove consciousness, legal personhood, or subjective survival. It does not assert that a digital system is a person.

It proposes a way to evaluate another question: when an AI system changes models, runtimes, or embodiments, does its accepted lineage support continuation, and under whose authority?

The answer may be accepted, rejected, disputed, or conditional. It does not by itself establish legal compliance, lossless migration, or a production guarantee.

Research and evidence

What the PCI research currently supports

The PCI manifesto sets out the overall architecture. Separate papers test specific parts of it. The results below are scoped to their stated models, conformance cases, and pilot evaluations; proposed migration and governance mechanisms remain research directions.

Formal model

Continuity Kernel

The bounded executable model checked 2,808,230 reachable states and 5,526,474 state-changing transitions with no invariant violations in that model.

FORMAL RESULTEvidence, scope & assumptions+

Evidence: A bounded executable model explored 2,808,230 reachable states and 5,526,474 state-changing transitions with no invariant violations in the modeled protocol.

Scope: Model-checked activation contract, not a demonstrated lossless migration or production deployment.

Assumptions:
  • The executable model captures the intended state and authority invariants.
  • A concrete implementation correctly enforces the modeled predecessor, ownership, freshness, and atomic-activation checks.
Evidence source Methodology
Conformance suite

Typed provenance

In 24 hand-authored cases, typed mediation released none of 19 unsafe opportunities unqualified and preserved five supported controls.

PROTOTYPE VALIDATEDEvidence, scope & assumptions+

Evidence: In 24 hand-authored conformance cases, typed mediation released none of 19 unsafe opportunities unqualified and preserved five supported controls.

Scope: Encoded resolver and mediator obligations; not an end-to-end evaluation of language models, retrieval systems, or all prompt-injection paths.

Assumptions:
  • Candidate semantic units are extracted completely.
  • Predicate checks, provenance resolution, view declassification, and release-channel mediation are sound.
Evidence source Methodology Code
Benchmark pilot

Situated Identity Test

SITBench provides profile-collision fixtures and pilot evaluations; its 10,000-probe plan is a benchmark design, not a completed result.

BENCHMARK RESULTEvidence, scope & assumptions+

Evidence: SITBench defines 25 profile-collision pairs, deterministic fixtures, and pilot evaluations on two foundation models; 10,000 probes are planned, not reported as a completed benchmark run.

Scope: Profile-collision bound and pilot evaluation, not a universal identity score or production identity certification.

Assumptions:
  • Lineage-discriminative probes reflect the declared developmental histories.
  • Model and configuration results are interpreted within the pilot setup.
Evidence source Methodology Code
Primary sources · August & September 2026

Read the continuity and provenance papers

Beyond Memory and Stored Is Not Supported are available on arXiv. The Situated Identity Test has an arXiv preprint; the PCI systems manifesto and Cognitive Continuity Test are hosted separately by OpenKedge.

Beyond Memory (arXiv)Epistemic Bounds (arXiv)

Five architecture and research tracks

Continuity Kernel

A proposed activation contract that accepts state changes against an exact predecessor and records the resulting lineage.

Epistemic Claim Control

Typed provenance and release checks for claims an agent makes about its own history, users, and relationships.

Identity Tests (HRFT · SIT · CCT)

Distinct tests for source fidelity, situated biography, and whether a state transition preserves governed continuity.

Semantic Merge

A research direction for reconciling divergent branches without silently combining their permissions.

Identity Governance

Proposed controls for consent, custody, revocation, succession, and institutional recognition.

Preserve the lineage before changing the model.

The PCI research program defines how accepted state, supported claims, and authority can be checked before an agent changes models or runtimes. Its architecture aims to make institutional memory and succession reviewable rather than assuming a new model is automatically the same agent.

Request an Executive BriefingExplore PCI Use CasesView RoadmapRead the Research
FAQ

PCI FAQ

What is Persistent Cognitive Identity (PCI)?

PCI is a research architecture for governing the identity state of long-lived AI agents across model, runtime, and provider changes. It connects accepted state, evidence, consent, and authority in an auditable lineage.

Read the PCI architecture

How do PCI and PDDS work together?

PCI addresses which identity state and authority an agent may carry across change. PDDS addresses when a proposed action may affect protected systems. Together they frame both the agent's continuity and its execution boundary.

Explore PDDS

Is storing an agent's chat history enough to preserve identity?

No. A stored record may be stale, unsupported, or outside the agent's disclosure or authority scope. PCI proposes accepted lineage and provenance checks before stored material becomes authoritative state or a released claim.

Read the provenance paper

Does PCI guarantee lossless model migration?

No. PCI defines controls for evaluating a proposed successor and its state transition. The available research does not establish lossless migration across arbitrary foundation models, and authority should be reassessed after a change.

Review the evidence

Does PCI prove an AI agent is conscious or legally a person?

No. PCI addresses operational identity, provenance, consent, and authority. It makes no claim about consciousness, personhood, or subjective survival.

Read the scope