A local, persistent memory layer. Scenes, decisions, outcomes — summarised, decayed, and surfaced as priors into the next reasoning pass. No cloud round-trip.
Storage lives on the edge node. No data exits the site unless explicitly opted-in.
Scenes, events, and outcomes indexed by time and by site.
SceneLM-distilled temporal summaries instead of unbounded frame history.
Tunable retention — events fade by importance, recency, and policy.
Hours, days, months — per event class, per site. GDPR-compatible.
Builds priors — what does Monday morning at this loading bay normally look like?
Replayable state. Audit-ready full envelopes for any past decision.
Opt-in cross-site federation. Share priors without moving raw data.
A camera that ran for six months on a real site is more accurate than the same camera on day one — not because the model retrained, but because the runtime accumulated context. That context lives in the Memory Engine.
The Memory Engine feeds priors into every reasoning pass.