An egocentric wearable camera is inherently invasive. It captures not just your life, but the lives of everyone you interact with. Major tech hardware treats this collateral data as a feature—uploading strangers' faces and voices to cloud servers to train future models.
Torres is built on a radically different foundation. We believe that an AI cannot be your true agent if it acts as a surveillance node for a third party. Privacy is not a toggle switch in a settings menu; it is the structural bedrock of the entire cognitive architecture.
The Data Sovereignty Standard
Every piece of data Torres perceives, synthesizes, and remembers stays entirely on your physical hardware.
Zero Cloud Telemetry
No audio, video, or episodic memory embeddings are ever transmitted to an external server. While the raw sensory perception layer executes via Unity Sentis and ONNX, the main cognitive stack is a completely custom, from-scratch model format built entirely in C# to guarantee secure, sovereign execution.
Local Vector Storage
Episodic memories are mathematically compressed and stored in local databases residing purely on your device's solid-state drive. If you lose your phone, the data dies with it.
No Third-Party API Hooks
Unlike wrapper applications, Torres does not rely on OpenAI, Anthropic, or Google APIs to parse reality. There are no API keys to leak and no user queries being silently logged by language model providers.
The Panopticon Contrast
To understand why building a completely custom local architecture is necessary, you have to look at the hostile reality of modern consumer tech. The internet has devolved into a massive, interconnected tracking apparatus, and wearable AI is poised to be its most invasive endpoint.
Telemetry & Shadow Profiling
Modern operating systems and smart devices operate as telemetry trojan horses. They constantly ping home, logging your DNS requests, location data, and daily habits to build cross-app behavioral profiles that are packaged and sold by data brokers.
The Cloud Processing Illusion
Tech giants claim that advanced AI computation must be handled in the cloud. This is a deliberate structural choice designed to force you to stream your raw, unencrypted daily life to their server clusters, turning your personal reality into free training data.
Collateral Harvesting
When a cloud-tethered wearable records a room, it completely bypasses the consent of everyone around you. It passively harvests the voices, faces, and conversations of your friends, family, and strangers to feed next-generation models.
Project Torres severs this pipeline completely. You cannot sell what you do not possess, and by running the stack entirely on local hardware, we act as a hard physical firewall against the surveillance economy.
Biometric Default-Deny
The Torres face and speaker recognition pipelines operate on a strict retention protocol. It actively categorizes incoming biometrics into known and unknown entities, applying aggressive pruning rules to anyone who isn't a recurring part of your life.
Unknown Entities
By default, people passing in the background or strangers speaking nearby are registered mathematically as transient noise. Their facial geometry and vocal imprints are utilized only for immediate situational awareness (e.g., "someone is walking past") and are aggressively pruned from short-term memory.
Organic Enrollment
Torres does not require awkward manual commands to learn a face. The system organically associates names when they are spoken naturally in conversation. If an unknown face is encountered more than once within a seven-day rolling window, Torres assumes they are an acquaintance and preserves the embedding. If they are not seen again within that week, the data is deleted. Rarely seen acquaintances can easily be re-learned dynamically upon their next appearance.
The "Forget Me" Protocol
Even after a person is enrolled, they retain ultimate sovereignty over their digital presence within your local system. At any time, an individual can speak the command "Torres, forget me" within the microphone's range. This triggers a specific security safeguard to solve the tracking paradox: while the face detector will continue to recognize that a physical face is present in the frame, the system is hard-coded never to pass that information up to the cognitive stack. If Torres simply deleted a facial embedding, the system would instantly re-perceive the person as a brand-new stranger a millisecond later. To prevent this, the Arbiter module intercepts the command, scrubs the person's name (converting their identity to a nameless hash), purges all historical episodic memories associated with them, and places a permanent hard-ban on saving any further data. In other words, though the perception layer can't help but detect forgotten people, Torres will be literally unaware that they exist. If you ask about a forgotten person, Torres will simply tell you that he did not see a person there, since he is consciously blind to their presence.
Transient Memory Pruning
During the nocturnal Dream State, the Osiris module runs its background training epochs to optimize the assistant's neural weights. As part of this routine, Osiris acts as a digital street sweeper.
Any peripheral visual data, unregistered voices, or irrelevant environmental noise from the day's events that did not contribute to a meaningful core memory is permanently destroyed. The system continually optimizes for signal over noise, intentionally discarding the massive, invasive datasets that traditional tech companies hoard.