est. 2000 — still learning
CHLOE
A persistent epistemic architecture for inspectable AI — separating durable cognitive state from a replaceable language-model interface.
About
CHLOE is a proposal for organising what persists behind an AI conversation. It separates a replaceable linguistic interface from a persistent epistemic substrate: the language model interprets and phrases natural language, but it is not Chloe’s memory, identity, or source of epistemic authority.
The persistent core stores explicit, scoped beliefs with provenance, distinguishes individual sources and domain-sensitive trust, retains unresolved questions across interactions, and revisits accumulated state during consolidation to surface contradictions and generate hypotheses for later verification. The current demonstrator instantiates a substantial subset of this architecture; richer long-duration agency and controlled self-modification remain research directions.
Chat
Talk to Chloe below. What you tell it is handled as attributed evidence: the persistent core records provenance, tracks trust by domain, surfaces contradictions, and can carry unresolved questions into later interactions.
?
The language model here reads what you say and phrases the replies, but it has no say in what Chloe believes. Small talk is answered and then forgotten; a claim is recorded as attributed evidence, with its source and a confidence that moves as other people agree or disagree; an input it cannot read confidently is refused rather than guessed at. What is being demonstrated is the belief-keeping underneath, not the conversation on top.
Try telling Chloe something, then contradicting that claim under a different name, then saying Sleep, Chloe.
History
- 2000CHLOE is first implemented in C++ as a seed architecture for persistent learning across interlocutors, with identity-conditioned admission of new material and sleep-time persistence.
- 2026The project is revived with modern language models as replaceable linguistic interfaces. A Python/SQLite prototype instantiates the epistemic core and powers this demonstrator.