Your work stays yours.
A personal store designed around what you create, with explicit access, durable saves, and a way back.
Your files are Memories.
Documents, books, photos, audio, films, game installations, and saved projects are presented as Memories. The model’s conversational context is a separate, explicitly labeled subset.
Object store underneath
Each Memory has a stable ID, typed content, ownership, timestamps, versions, provenance, and access grants. Binary blobs live in a persistent filesystem-backed store; a transactional metadata index handles names, relationships, and search. Embeddings are derived indexes that can be rebuilt.
Use separate classes for personal creations, installed software assets, game saves, task history, and AI preferences. This keeps “uninstall game” distinct from “delete my saves.”
A bridge to existing apps
The Memory broker exposes selected objects as ordinary paths inside each Pod. Save operations create versions. Mounts can be read-only, read-write, or one-time imports; apps cannot browse all Memories by default.
Provide export to normal directories, encrypted local backup and restore, explicit deletion with retention controls, and key recovery choices. Persist the metadata and content before marking a save complete.
A Memory has a history.
- Create.
A new manuscript receives a stable identity. Its name and location can change without changing what it is.
- Save.
Content and metadata must be persisted before the system confirms success. A new version records a change you can understand.
- Share access.
A selected Pod receives only the objects and operations you grant. A one-time import does not have to become a permanent mount.
- Restore or export.
Use version history, backups, and ordinary directory exports to carry your work forward. Recovery and key choices belong in the product experience.
Remembering your work
is different from learning about you.
A document is not automatically model context. NIX separates your creations, software assets, saves, task history, and AI preferences. The architecture should make it clear which data a task can read and which preferences the model can use.
Try Author and Memories ↗