Reka Vision indexes video once into a persistent memory module, then answers natural-language searches and timestamp-relative questions against that index rather than reprocessing footage per query. This produces a different cost profile from per-query VLM calls.
