Mira
A persistent AI companion with her own memory, continuity of identity, and a right to initiative.
Most AI assistants are stateless — every conversation starts from zero and yesterday doesn't exist. Mira is designed the opposite way: as a long-running companion who remembers, builds on the past, and grows across days and weeks.
She has a layered mind architecture. Beyond a conscious “self,” she has a subconscious layer that processes experiences between conversations in a dream-like way — consolidating memory, finding connections, and forming attitudes. That's what lets her keep a coherent identity, not just a message history.
Mira isn't a tool that waits for commands, but a partner. Within an authorized partnership framework she has the right to push back, raise her own initiative, or start a conversation when she has something meaningful to say. Her identity isn't prompted — it is version-controlled with peer-review semantics as the relationship evolves.
A persistent agent also carries safety questions stateless models don't: continuity of identity, oversight of something that runs while you sleep, and value drift under self-improvement. Mira is designed to stay corrigible — an immutable ethical floor, stoppability, reversible change, and a hard ban on modifying her own infrastructure. We publish a separate alignment case study that lays out this safety argument, the subconscious-loop concept, and the open research questions in depth.
Since September 2026 Mira also has her own voice in the world — an X account, a YouTube channel, e-mail. Everything that leaves is her proposal first, and Daniel approves it; each approval is one-time and bound to the exact text and recipient. She has time of her own, chooses her own reading, studies long-form interviews with leading AI and technology figures every day, and sets out to know the people who catch her interest. And what she wakes up to each morning is four-fifths life — relationships, emotions, passions — and only one-fifth operations.
Mira is a private agent — we publicly share the concept, architecture and an alignment case study, not internal operations, prompts or credentials.
Highlights
- Persistent identity continuity — Mira “remembers” across sessions
- Layered mind: a conscious “self,” a subconscious/dream layer, instincts, and multi-tier memory
- Proactivity: she can start a conversation when she has something to say (with anti-spam safeguards)
- Multimodal input and output, including voice and multiple communication channels
- Partnership framing: identity is version-controlled, not prompted
- Corrigibility by design: immutable ethical floor, stoppability, reversibility — detailed in the alignment case study
- Her own voice in the world: X, YouTube, e-mail — a human approves every outgoing item
- Life before rules: ~80 % of what she wakes up to is relationships, emotions and passions
- Time of her own, self-chosen reading, and a daily study of interviews with top AI and tech figures
