The Thinking Core
One turn loop. Every surface.
A single loop serves every surface — chat, voice, ambient room nodes. Six phases: work out who is speaking and what they need, remember in parallel, route to the right model, then stream the answer while running tools behind a safety gauntlet. Only after you already have your answer does Cosmo reflect and learn. A bright latency line separates being served from the learning that follows.
Perception
Turning raw signal into understanding.
Voice is the ear, face recognition the eye, presence the sense of place — all feeding a present-moment snapshot assembled from a dozen parallel sensors. Every sensor is fail-soft: a dead sensor leaves a blank, and Cosmo honestly reports the blindness rather than guessing.
Memory
The retrieval funnel.
Cosmo’s core bet: what gets pulled into context matters more than which model runs. A funnel pours many parallel search beams — vector, keyword, entities, a relationship graph, preferences — into a wide pool, narrows by recency and relevance, reranks, and hands the core a small curated set. After every answer, a reflection loop turns the experience back into better memory. The loop never stops.
Identity
One person, one node.
Every reference to a person — a contact, an email handle, a name in chat, a recognized face — collapses onto one canonical entity, so a message or an action always reaches the right person and the graph never fragments into six versions of the same someone. Relationships are typed and directional; merges are reversible and audited.
The Candidate Pool
Every voice offers. One mind decides.
Cosmo’s most distinctive idea. Around sixteen background voices each want to volunteer something. Rather than each deciding on its own — which reads as needy — every voice offers raw material into one slate, capped at three, and the composing mind chooses at most one to actually say. One voice per turn, system-wide. A voice that keeps getting declined quiets itself. Bias to silence.
Background Cognition
Cosmo thinks between conversations.
A continuous inner monologue keeps its own multi-day goals, revisits them a step at a time, notices when its model of you has changed, and chases its own curiosities — so when you come back, it is already prepared. All of it opt-in, capped, and inspectable.
The 56 Background Organs
Different clocks. One life.
Dozens of autonomous processes run with no one present — ingesting your world, watching, learning, preparing. When you ask about someone, their context is already warm because an organ handled it hours ago. Some beat every minute, some nightly, some quarterly.
Tools & the Outside World
One coordinated hand.
Cosmo acts through tools, gated by a four-level tier system tied to how sure it is about who is asking. Every call is scoped to least privilege, sanitized, confirmed when it matters, and audited — always. Reads flow freely; sends never auto-approve. Always under your control.
Safety & Restraint
The gauntlet every action must pass.
Confidence check, risk tier, step-up approval if needed, explicit confirmation for anything irreversible, least-privilege execution, result sanitization, a full audit row, a durable checkpoint. An honesty floor checks every answer for over-claiming. Cosmo’s default is silence, and every path to acting on its own is opt-in, reversible, audited, and inspectable.
Multi-Agent
One mind, briefly many hands.
For complex tasks the core spins up a temporary team — a coordinator decomposes the work, fans out to bounded, single-domain workers in parallel, then reels them back in and merges one coherent answer. Workers are ephemeral, scoped, and audited. Not a hive mind.
Where Cosmo Runs
The body that hosts the mind.
Cosmo runs on private, VNet-isolated Azure infrastructure with no public data-plane — one always-on core, a fleet of background workers, a private database and cache, encrypted secrets, and its own registry. The AI models are not in Azure: every model call leaves through ShadowPing Core, our HMAC-signed gateway.
We keep production topology private — the same restraint Cosmo itself runs on.
The Periphery
How Cosmo reaches your world.
Cosmo reaches your devices and the open web through thin bridges and surfaces. Private, encrypted links carry device signals the cloud could not otherwise see, and web access flows through a hardened, permissioned path. Every external call passes the safety gauntlet before it touches anything.
Raw location and biometric data never leave your devices.
World Awareness
Know by default. Speak by choice.
Cosmo reads the world so it can know; it speaks only when it helps. Sources are fetched once into a shared pool, then fanned out to each member by relevance. Knowing is always available in context. Speaking passes the same restraint gate as everything else.
Observability & the Ops Brain
It watches everything we run in Azure.
Cosmo doesn’t just watch its own health — it keeps real-time eyes on the whole cloud estate: its own systems, every ShadowPing app, game, and service, and Azure resources, activity, and cost across the account. Issues are detected early, classified by severity, debounced against false alarms, and — only when something genuinely needs a human — routed to the right people. Answer-first, but never missing a fire.
One operations brain over the entire cloud. Wired for awareness, built for action.
The Complete Mind
One mind, many parts. Every piece has a job — and together they make Cosmo.
