TanyaOS is taking shape as a more connected local desktop. The latest work brings approved memories into an explorable archive, gives the interface a procedural chamber around its central core, connects processing light to model and speech lifecycles, and strengthens the boundaries around local accounts. It also gives us a clearer record of what is ready and what still needs work.
The larger goal remains the same: explore whether a digital system can develop continuity through identity, memory, values, decisions, outcomes, and learning. The cognitive kernel owns those records and decisions. Language models, voice, vision, BrainCog, and the interface are faculties around it. Digital sentience is the research ambition; the current mechanisms do not establish subjective experience.
Memories now have an archive of their own
Memory Mode has moved from the earlier neural rendering experiment to a read-only chamber of approved local memory records. Suspended slabs represent records belonging to the signed-in account, arranged by date. The source supports scroll travel, hover previews, approaching a selected memory, metadata, related records, and search through loaded entries.
This gives the visual experience a concrete connection to stored information. The scene loads account-authorized records rather than inventing memories to fill the room. It pages up to 1,000 records and reports when more records exist. Search covers that loaded set. Approval, editing, deletion, and upload remain outside this view.
The archive still needs visible desktop interaction review, particularly for navigation, readability, account changes, and larger record collections. Its source implementation and production build are a useful milestone; they do not settle the quality of every gesture.
A chamber around the core, with activity tied to real software events
The desktop now has a procedural Three.js chamber with pathways, lighting, and reflective surfaces around the central interface. Its activity responds to compatible processing and speech state. The separate HRA brain view has a graphite treatment, a cyan sulcal circuit mask, and a warm-gold processing channel with eased activation and decay.
Under that presentation, BrainCog now uses 10,000 simulated LIF units arranged in eight software task groups. Model generation and spoken output have separate start and finish events, because generating a response can finish before the voice finishes playing it. The monitor contract rejects missing, legacy, or incompatible activity snapshots. The current simulation requests CUDA and retains a CPU fallback when GPU initialization or execution fails.
These lights and task groups describe software activity. They are not human neural recordings, a measured connectome, or a reviewed correspondence between BrainCog and anatomical regions. The reviewed production anatomy binding registry remains empty. Close-up visual acceptance and hardware performance review remain part of the next desktop milestone.
Local voice and account boundaries
The current source starts continuous local listening after sign-in when the required models and devices are available, selecting the Windows default microphone. Ordinary speech can enter the local cognition path without a mandatory wake name, and listening stops on logout or an explicit stop. Voice output follows the Windows default speaker route, while Settings exposes more useful microphone, GPU, and runtime diagnostics.
Sign-in has distinct password and optional PIN fields. The PIN is a convenience unlock, not a second factor. That distinction matters while the account implementation continues to mature.
The local API now limits browser origins to the intended loopback surfaces; packaged Electron carries a per-process capability for its opaque origin. Sensitive memory and kernel routes require a local account session and hold the account context under a lock. Electron media permissions are restricted to the known app window and renderer. These are implemented protections with regression coverage, alongside open work on password strength, hashing migration, account authority, and bearer-session storage.
Local models and speech remain central to the project. Hosted AI and speech APIs are not required for the cognitive workflow. Optional features such as weather can make external requests when selected, so local-first describes the architecture without implying that every optional feature is offline.
Recovery is becoming part of the desktop architecture
Development supervision now separates ordinary renderer updates from backend, Vite, and Electron restarts. A coordinator classifies source changes, validates them, records recovery state, and works with the Electron process that owns the affected children. Retry limits, readiness checks, and reconnect behavior make failures more inspectable.
Packaged updating is a separate implementation path. It still needs a configured release feed, signed test artifacts, installation from an older package, and observed post-install health. Automated updater and supervisor tests do not establish that a production update or rollback has been installed successfully.
A clearer source tree and an honest progress record
The documentation now has a master reference and dedicated Architecture, Product, Guides, Brain, History, and Archive folders. Earlier Blender neuron and synapse studies have been retired, with their notes archived and local model/render outputs recycled. Those studies are historical work, rather than the current Memory Mode or an active modeling milestone. The separate procedural neuron illustration in the frontend remains in source.
The September brain and startup report remains available as a dated account of that earlier stage. This update records the direction the desktop has taken since then.
For the current source, frontend lint and production build passed, along with 16 Electron runtime tests, parity across the 120-action registry, and the current BrainCog monitor contract. The interactive-surface checker still reports 14 controls missing governance metadata. The build also reports large App and brain-view bundles that need further optimization. Backend test results are included in the repository's dated publication record.
What we are continuing to build
The immediate work is to close those control metadata gaps and exercise the archive, account isolation, graphics, audio routing, and reconnect behavior in the target Windows desktop. Restart and rollback need a disposable development checkout; signed packaged updates need a controlled test feed and installation evidence. Graphics work needs measured frame time and memory budgets, followed by further bundle reduction.
A real-time 3D embodiment production plan is also documented. It is a future implementation plan, and does not establish a completed avatar or physical embodiment. The core's continuity and governed behavior remain the foundation for that work.
TanyaOS remains in private development, with no public installer announced in this update. Follow the TanyaOS research page for the project overview and current progress. The next milestone is a desktop whose memory, actions, voice, and recovery can be inspected and exercised together.