{ COMPUTER VISION / MEMORY GRAPH / OPEN SOURCE }Live
Intelligence OS.
Turns camera streams into a memory you can question in plain English — not footage you have to scrub.
In-house · Infinex Labs
// Overview
What we built
An open-source, self-hosted system that watches camera streams, spends compute only when something changes, and writes what it saw into a persistent memory graph — who and what appeared, where, and when. The same person stays one entity across days and cameras, every claim traces back to a keyframe, and the memory can be named, merged, split, or deleted. Ask it "was anyone at the loading bay after six?" and get a grounded answer with evidence.
// Specs
Key facts
Pipeline
Nine-stage cascade — cheap gates before paid VLM calls
Memory
Persistent entity graph with frame-level provenance
Deployment
Self-hosted, works fully offline — no cloud keys required
License
Open source, Apache-2.0
// Features
What it does
01
Natural-language questions answered from stored observations, not model guesses
02
Cross-camera identity that keeps one person as one entity across days
03
Zones and rules compiled from English, with unanswerable rules refused at compile time
04
Habits and daily digests that surface what broke pattern
05
Correctable memory — name, merge, split, cascade-delete
06
Privacy by design: opt-in face recognition, keyframes not video, no audio, no telemetry
// Architecture
How it works
Flow
Camera Streams
Motion Gate
Detection + Tracking
Identity / Re-ID
Scene State & VLM Description
Memory Graph
Distillation (Habits & Events)
Ask / Operator
Interested in
Intelligence OS?
If you have a problem shaped like this one, tell us about it. Most of our work starts from a short email.
Get in touch→// End of case studyIntelligence OS