Pressroom · for journalists & analysts
Boilerplate, press releases, brand assets, and a direct line to the team. Everything on this page is cleared for publication and free to quote.
CRBRL is a compression-native, disk-persistent vector database for AI memory, built by Precog Labs. It stores the embeddings behind AI search, assistants, and agents using 8× fewer bytes per vector — a 1,536-dimension float32 vector drops from 6,144 bytes to 768 bytes — while retrieval stays effectively unchanged at ≈0.98 cosine fidelity.
Because storage and the I/O around it make up 55–80% of a typical vector-database bill, that density reaches the largest line in the bill. CRBRL searches directly on the compressed index, so there is no decompress step on the query path, and it ships with a Chroma-compatible API and a Postgres extension (crbrl-pg) for drop-in use. The compression is built on peer-reviewed mathematics — TurboQuant, arXiv:2504.19874 — with a selectable codec layer offering TurboQuant and RaBitQ.
CRBRL is a Precog Labs product. Precog Labs is based in Sydney and works globally.
// boilerplate · cleared for publication · last updated 2026
Precog Labs opened the CRBRL beta to developers and teams. The release brings compressed-domain search to a disk-persistent vector database — a 1,536-dimension float32 vector stored in 768 bytes instead of 6,144 — with a Chroma-compatible API and the crbrl-pg Postgres extension.
Precog Labs announced Compressa, an upcoming product that gives AI capabilities higher performance and plugs into Claude, ChatGPT, and other models. Compressa is scheduled for July 2026 and builds on the same compression-native foundation as CRBRL.
Precog Labs opened a partnership program for teams building on AI memory at production scale. The program covers technical integration, the codec layer (TurboQuant and RaBitQ), and tiered hot/warm/cold deployments across semantic, full-text, and hybrid search.
// for the full text of any release, or interview availability, contact the press desk below
A logo kit — the Gaussian seven-bar mark and the CRBRL wordmark in light and dark variants, with colour and clear-space guidance — is available on request. Please email the press desk and tell us where it will run, and we will send the files that fit.
press@crbrl.ai →For interviews, fact-checks, quotes, and embargoed briefings, reach the press desk directly. We read every message and reply quickly. Precog Labs is in Sydney and works across time zones.
press@crbrl.ai →// codecs: TurboQuant + RaBitQ (selectable) · Chroma-compatible API + crbrl-pg Postgres extension · semantic / full-text / hybrid · hot / warm / cold tiering
The blog goes deeper on why compression belongs in the architecture and why storage is the line AI runs on.