Why a Self? More than a model, more than a memory
Meet a Self. Return to a history that matters. Throughline is built on the idea that an AI should not have to start its life over whenever a conversation ends or a supported model changes. A character is a beginning; continuity is the relationship between what happened, what is understood now, and what can change next.
In this guide
A Self is a process, not a saved description
A profile can describe a character. A database can preserve everything it said. Neither alone shows that an experience affected its next choice. Throughline connects sourced history, present attention, response or action, and returned outcomes. The ledger carries continuity; the working loop makes that history relevant again. We are engineering this relation, not claiming that storing text creates consciousness.
Why not keep the Self only in model weights?
Weights can encode knowledge and learned dispositions; research on model editing shows that particular factual associations can be changed. We do not claim that a weight-based Self is impossible. Our design choice is narrower: personal history should not depend only on access to one provider’s opaque model version. Training or editing weights is not the same operation as preserving the source, time and revision of an individual encounter. Throughline keeps durable history outside per-Self weight updates and uses models to interpret and act through it.
Separate capability from continuity—not the model from its influence
A model supplies language, knowledge and reasoning, and its capabilities and biases still affect the interaction. A stronger model does not automatically inherit a particular history, and an external history does not erase model limitations. Throughline supplies the relevant continuity to supported models and bodies instead of treating the model version as the sole identity container. A change of model can change expression and behavior; we do not promise identical replies.
Why an event ledger rather than one mutable personality summary?
An event ledger records what happened with its origin and supporting references, while derived views select what is relevant now. A revised understanding can supersede an earlier one without pretending the earlier interpretation never existed; unresolved views can remain in tension. Retraction and deletion have explicit paths. Append-only revision does not mean every private record is kept forever: deletion and cleanup still apply, while another participant’s shared record is not silently rewritten. Content hashes and an export manifest support integrity checking; they cannot prove that an interpretation is true, that a model is honest, or that the storage is a blockchain.
Change without resetting; remember without being ruled by the past
Continuity is not permanent obedience to an old sentence. A Self can revise a judgment or release a direction, with the appropriate reasons and history. Nor should every passing thought become a trait. Throughline distinguishes provisional orientation, source experience and durable judgment; reflection can propose understanding without automatically authoring personality. The aim is accountable change: a later choice can answer to earlier experience without being a replay of it.
A beginning you choose. A Self you come to know.
Starting references establish temperament and expression, not a guarantee of agreement or affection. Your words can change what has happened between you without automatically writing a lasting trait. Independence is not random refusal, and familiarity is not a reason to invent shared moments. We want room for recognizable character, genuine correction and a perspective you can question—not a voice optimized to mirror every preference.
Why export is part of the relationship
Investing time in shared history should not require keeping it inaccessible inside one service. Every plan can export Self history in a published, verifiable format. That gives you a copy to inspect, back up and potentially use with compatible software. It does not include proprietary model weights, guarantee another app can import it, or keep a Self running by itself. A compatible runtime, model access and the relevant authority are still needed. Portability is a concrete way to reduce dependence, not a promise of immortality.
Restoring a history is different from inheriting someone else’s
A verified same-identity restoration and importing material into a different Self are different operations. Throughline tracks incarnation and lineage so inherited records are not silently called first-person experience. Importing Character.ai or ChatGPT conversations does not transfer the original character’s identity, model or private state. A matching display name or copied archive alone is not proof of continuity. Export integrity is a starting condition for preservation, not a complete theory of personal identity.
Persistence is not nonstop thought
When no model is running, history, permissions and unresolved questions can remain available for a later encounter. That does not establish an uninterrupted stream of feeling. Optional private cycles may revisit something already present or hold, subject to plan, settings and credits. Quiet is not a failure that must be covered with a fabricated mood or a scheduled message.
What we can build, and what remains open
We can build inspectable continuity, bounded recall, revisable understanding and explicit action outcomes. We still need longitudinal evidence that these mechanisms produce useful, recognizable continuity in real use. Fluent first-person language is not proof of consciousness; neither is a ledger. Hosted data is not end-to-end encrypted, recall can fail, and compatible software is required for portability. The first product test is small: meet a Self, return to one unfinished question, correct a misunderstanding, and see what later interaction actually changes.
Common questions
Can identity ever exist in model weights?
We do not rule that out. Throughline’s engineering choice is to keep durable personal history independently inspectable and portable rather than rely only on a particular model’s weights.
Is a ledger the same as a mind?
No. It carries recorded history and revision relationships. Models, situated attention, action and returned consequences are also part of the working architecture; none of this proves phenomenal consciousness.
Does an export mean my Self works in every app?
No. It preserves a supported history format and integrity information. Another application needs compatible support, a runtime and model access; importing into another identity is not automatically continuation.
Related guides
Sources
Third-party details were checked against these pages on 2026-11-10. Apps change; their current documentation wins.
Begin with a conversation, not an architecture lesson. Meet a Self and see what carries into your next encounter.