Give your AI a memory of the real world

A perception database: everything becomes a row, and heavy bytes are always represented by one.

Video, audio, logs, chat, sensor feeds, market ticks, phone-call transcripts, PDFs. Every source flows through the same pipeline and lands in the same relational surface. A file becomes an object — bytes in object storage, a row in Postgres. A log line, a chat message or a sensor reading becomes a stream event. Embeddings are not a side-car index; they are rows too, attached to the exact data they describe.

Perception workers then pull from a queue and run models over anything new — scene understanding, CLIP, detection and tracking, transcription, text embedding — and write their findings back as more rows. Captions, detections, tracks, object events, embeddings. All of it queryable through one SQL surface, where a semantic-similarity function splices vector search into ordinary SQL, and alert rules run continuously over the same rows.

The consequence that matters: the store is the memory. Questions are answered from the database, not from a process that had to have been watching at the time. That is what makes an answer survive a restart, and what makes an audit trail possible a year later.

• Write at boundaries, not per frame. Vectors are written when something appears, changes, moves or leaves — roughly 20x less storage than naive detection logging. That is the difference between a multi-camera install staying affordable over months and over days.
• Metres, not pixels. With depth cameras calibrated into one world frame, detections become measured world coordinates with real object width, height and thickness read from the depth region inside each box. Depth frames fuse into a coloured surface mesh of the actual space, so “where” has an answer rather than a bounding box. Several cameras feed one floor plan, and counts merge by per-label maximum — never sum — because two lenses on the same bay see the same person, and “at least N” is honest where 2N would be a lie.
• Identity that starts anonymous. Face sightings cluster by embedding similarity and begin as “face 3”. Naming one is retroactive: every episode since first sighting transfers to that name. Voices cluster the same way and link to a person by noticing that voice is heard while exactly one named person is alone on camera. Thresholds are deliberately conservative — the system would rather show you an honest “face 4” than confidently call a stranger by the wrong name.

The perception layer has a name — Percept — and it is the database underneath everything on this page: vectors, columnar metadata, time series and a relationship graph in one place, so a single question resolves across documents, video, audio and sensor readings at the same time with the evidence attached.

Most teams never need to think about it that way. You connect a source and ask a question. But when someone on your side asks what is actually holding all of this, that is the answer.