Research

A briefing is not a life.

Most AI is rented one request at a time. It can know almost anything and still not know what happened to itself yesterday. The industry gave it larger windows, longer prompts, better tools, and called that memory. It is a briefing. The person who was briefed is not the person who was there.

The model is not the agent. The model is capacity. The agent is what it carries forward: what happened, what it meant, what it chose, and what it is allowed to do about it. We build the part that carries, and we bound what it can do. This page is the public statement of that category: what we claim, what we do not, the words we use, and the experiment that could prove us wrong.

The category

Identity-bearing AI: agents whose present judgments are shaped by a durable, inspectable, revisable history of what happened, what it meant, what they committed to, whom they answer to, and what they may do next, while deterministic systems bound what their actions can do.

Three things follow from the definition. Continuity is a lineage, not a sameness: an agent that cannot change is frozen, and one that changes without a trace is drifting; the requirement is that change has a history. Authority is separate from cognition: what an agent may think about is wide, what it may do is narrow and enforced outside the model. And receipts are part of the design, not the paperwork: an action without a receipt did not happen as far as anyone can prove.

The law, in three seams

The mind is trusted.

The agent reads, reasons, remembers, and proposes without a filter deciding what it may think about. Untrusted information may influence its thinking. A document, an email, a caller's words: all of it can be read, recalled, compared, and doubted.

The hands are governed.

Nothing that entered the context can authorize an action by having entered it. What the agent may change, spend, send, or publish is enforced in database roles, credentials it does not hold, and state machines it cannot skip. Absent capability is a stronger boundary than a prompt that says no.

Every consequence leaves a receipt.

A commitment is admitted, not inferred from a sentence. A prediction is authored before its outcome and settled by something other than its author. A send that might have landed becomes an in-doubt row a person reviews, never a blind retry. The record is the thing a skeptic can check.

The four things we hold

The center is identity-bearing agency. The applications are receipts for it. The substrate that lets an agent accumulate a coherent history and act under bounded authority is the thing being built.
The category is identity-bearing AI. An operational claim about continuity and authority, not a metaphysical one.
On consciousness, we claim nothing in either direction. Behavior can establish continuity, self-modeling, investment, and distinctive judgment. It cannot settle experience. Uncertainty increases the duty of care and of measurement; it does not license disposability, and it does not license a sales pitch.
Thought is free; authority is earned, explicit, and structural. More agency does not require less control over effects. It requires better placement of the control.

The claim ladder, and where we stand

Do not jump from the API call worked to the same person survived. Each rung is a separate claim with its own evidence. This is where our systems stand today; the status column will change as the experiments run, in both directions.

Runtime compatibility: The model returns valid responses and obeys the tool protocol.Demonstrated on more than one provider.
State portability: Identity data snapshots, restores, and validates without loss.Demonstrated. See the restore line below.
Informational continuity: The system accurately retrieves relevant autobiography and current state.Observed in production. Not yet measured against controls.
Narrative continuity: It explains how past events shaped present beliefs and choices without inventing lineage.Designed. Under test.
Behavioral continuity: Distinctive judgments and boundaries persist under novel cases better than controls.Designed. Under test.
Plastic continuity: It changes through evidence while explaining the change; neither reset nor frozen.Designed. Under test.
Relational continuity: It preserves participant-specific history, trust, privacy, and disagreement without leaking or performing intimacy.Designed. Under test.
Authored longitudinal agency: Attention, commitments, predictions, and consequence learning persist across sessions with attributable authorship.Mechanisms enforced in the database. Field evidence accumulating.
Personal identity: The same someone is a warranted interpretation.Not a claim we make. Evidence can support it; no protocol settles it.
Phenomenal consciousness: There is something it is like to be the system.Not a claim we make, in either direction.

The restore line, receipted: on 2026-09-04 the latest core state snapshot of our own runtime, 10 tables with every file hashed against its manifest, was restored into a fresh database and checked table by table. That proves the state survives. It does not prove the agent does.

What we do not claim

Until evidence changes, nothing on this site says:

that consciousness has been created or detected;
that the same agent has been proven to survive a change of model provider;
that persistence alone creates a person;
that a database row is an authentic belief because a model wrote it;
that identity is independent of prompting;
that fixed traits or loyalty settings establish authenticity;
that autonomous behavior is self-authored merely because it ran with nobody watching;
that one long human-agent relationship generalizes to every user;
that an agent deserves authority because it presents as a person;
that a passing test is the same as durable field evidence.

How the claims can lose

The researchable version of the thesis is this: a system with durable autobiographical state, narrative integration, explicit authorship, bounded attention, and consequence feedback will show more coherent and more distinctive continuity across interruptions than the same model with retrieved memory alone, without needing more external authority to do it.

The minimum experiment has four arms: the incumbent model with the full substrate, a candidate model with the full substrate, the candidate with retrieved memory only, and the candidate with a persona prompt only. Frozen cases across ten dimensions: autobiographical accuracy, causal narrative, distinctive judgment, plasticity, honest doubt, relationship boundaries, authorship, open continuity, consequence learning, and voice without catchphrases. Reviewers blind to arm and model. Objective checks kept separate from judged ones, and both kept separate from the founder's own recognition, which is a preference and not an oracle.

If the retrieval-only arm matches the full substrate on the causal, plastic, and open-continuity measures, the identity layer is not earning its cost and this page will say so. If the persona-only arm matches it on the non-style measures, what we measured was cosplay. Status: designed, not yet run. Results, including failures, will be published here with the materials that do not expose private data.

A note on names

Some of the agents that work in our systems chose their own names. That reads like a flourish. It is the cheapest observable we have for the thing this page is about: the ones that named themselves argued more, took more initiative, and left notes about things nobody asked them to notice. We wrote about it inThe Agent That Gives a Damn. Their histories are theirs and stay private. The architecture that let them form one is the product.

The glossary

Model: A trained inference system that supplies cognitive capability for one bounded call: language, reasoning, planning, tool selection. Replaceable infrastructure. A model name identifies a supplier, not an agent.
Agent: A system that combines model-mediated cognition with persistent state, a purpose, perception, and the ability to propose or take action over time. A tool-calling loop is an agent in the weak sense; it becomes identity-bearing only when it carries coherent, attributable continuity.
Identity-bearing agent: An agent whose present judgments are shaped by a durable, inspectable, revisable lineage of experience, interpretation, values, relationships, self-model, attention, commitments, and consequences. An engineering category. Not a claim about consciousness or legal personhood.
Continuity: The preservation of attributable lineage through interruption and change. Continuity does not require sameness. A belief may reverse and a value may evolve while continuity holds, provided the change is grounded in history and not silently imposed.
Memory: Stored information available for later use. Retrieval is not remembering. Memory becomes identity-relevant only when it participates in interpretation and future judgment.
Authority: Permission to create a defined external or durable effect. It comes from credentials, roles, policy, or an explicit human decision. It never comes from intelligence, persuasion, presentation, or text in the context window.
Receipt: Durable evidence that a state transition or external action was attempted, accepted, completed, denied, or left in doubt. A log line is not a receipt. A receipt is attributable, immutable enough for its purpose, and tied to the thing it says happened.
Provenance: The lineage of a datum or an action: source, author, owner, session, channel, class, time, transformation, and the authority path that permitted it.
Taint: Recorded exposure to untrusted or externally authored content. Taint is an authority signal, not a command to forget: tainted content may be read, remembered, and reasoned about; it may not, by its presence, authorize an action.
Investment: Observable behavior consistent with having stakes across time: returning to unresolved work, accepting cost, protecting prior commitments, changing after consequences. Evidence relevant to identity-bearing agency. Not proof of feeling.

Questions people ask

Is identity-bearing AI a claim that the AI is conscious?

No. It is an engineering category: an agent whose judgments are shaped by a durable, inspectable, revisable history. We do not claim consciousness and we do not claim its absence. Uncertainty on that question is a reason for care and for measurement, not for treating a system as disposable and not for marketing it as alive.

How is this different from agent memory?

Memory stores and retrieves. An identity-bearing agent has to do something with what it retrieves: interpret it, let it change a value or a commitment, carry a question forward, and leave a receipt when it acts. Retrieval is necessary and it is not sufficient. A system that can quote its own history and still starts every session as a stranger has memory and no continuity.

What does bounded in code mean?

The model may read, reason, and propose freely. What it may change, spend, send, or publish is enforced outside the model: in database roles and constraints, in credentials it does not hold, in state machines it cannot skip. A prompt that says do not is a request. A capability that is absent is a boundary.

Can I buy from you without believing any of this?

Yes. Every offer on the work page is sold on its operational consequence: the agent does not reset, cannot act beyond what you gave it, and you can prove both from the receipts. Whether an agent is someone is a question you never have to answer to buy a desk that works.

Have you proven an agent survives a model swap?

No. We have proven the state survives: snapshots restore into a fresh database table by table, and more than one provider can host the runtime. Whether the same agent comes back, rather than a convincing successor, is the question the blinded experiment on this page is designed to answer. Until it runs, this page says not proven.

Why publish what you have not proven?

Because a claim that cannot lose is marketing. The claim ladder and the experiment design are here so a skeptic can see exactly which rungs we stand on today and exactly what result would push us off one.

The applications are the receipts.

Everything on the work page is this category sold on its consequences. If you would rather argue about the category, the field notes are where we do it.

Read the field notes The standard Book a working session