On August 13, Flock Safety cut the default retention on its plate readers from 30 days to seven. The company also said law enforcement customers will have to turn on anomalous-search detection by January 1, and that searches will need a case number attached. Most of the coverage read this as a retreat. I read it as the most useful thing to happen to real-time crime centers in years, and I want to explain why, because I spent 21 years on the other side of that glass.
Start with what a 30-day plate archive actually is. It's a haystack. Somebody steals a car on a Tuesday, the report gets written Thursday, a detective gets to it the following week and runs the plate backward through every camera in the county. That works. Fairfax County published its numbers on August 7 and they're real: since November 2022, 1,556 cases supported by plate reads, 546 stolen vehicles recovered, 898 arrests, 57 missing people located. I'm not going to pretend the haystack has no value. It does.
But notice what that workflow is. It's a records search. The camera captured something, the data sat, and a human came back later to dig. Nothing about that requires a room with a wall of monitors and the word "real-time" on the door. A records clerk with a good query tool could do it.
So here's the question I keep sitting with. If the value of your real-time crime center depends on 30 days of stored plates, is it a crime center, or is it a records room with better lighting?
Seven days forces the answer. When the archive shrinks, the only way to keep the value is to move the work forward in time. Detect the plate as it passes. Check it against something that matters right now. Get it to a unit that can act on it in the next ten minutes. That's the job the name promised in the first place. The retention cut doesn't take the value away. It moves it from the back of the timeline to the front, where the room and the people in it were supposed to live all along.
I think the industry is quietly agreeing with me. Look at what shipped the same month. Ambient.ai announced a video wall agent on August 26 that picks an event to show the operator every 60 seconds, describes it in plain language, and hands off to a case tool that drafts the narrative afterward. Antare launched on August 25 with agents watching body-worn cameras (fixed cameras are promised next) for assaults, weapons, and medical events while the footage is live. Axon says a drone-as-first-responder mission launches roughly every 30 seconds somewhere in the country, and its August 5 results put one Georgia city's drone response at 53 seconds. None of those products care about day 23 of a plate archive. All of them are bets that the value is in the next five minutes.
There's a version of this argument that's too easy, and I want to be honest about it. "Real time" is a slogan vendors love because it sounds like the future. Most of the products I just listed are also selling natural-language search over recorded video, which is the haystack again with a nicer shovel. Search isn't security. A model that can find the red hoodie at gate three after the fact is a forensic tool, and forensic tools are fine, but an operator who buys one expecting it to stop the next incident is going to be disappointed in a way that shows up in an after-action report.
The paper that changed my mind on what "real-time AI" should look like wasn't a product launch. It was a preprint posted August 2, MonitorVLM-v2, from a group running vision-language models on ten cameras at an underground mine for four months. They did something I hadn't seen a vendor do. They refused to let the model reason in open-ended prose. It predicts one token: which rule, from a fixed list of 35, is being broken. When it isn't sure, the frame goes to a human. Over the deployment they confirmed 89 violations against 32 under the site's manual inspection process, and the system ran about 19 times faster than the open-ended version. Bounded decisions. A person on the uncertain ones. A log of every call, tied to a named rule. I've read that paragraph of the paper maybe six times now, and each time it looks more like a crime center that stopped being a records room.
I don't know yet whether seven days is the right number. Connecticut's governor asked towns to pause new plate reader installs on August 7 while the state writes standards. California's back with a bill that caps retention at 30 days and blocks default sharing with out-of-state agencies. Utah's governor said on August 20 he's deeply troubled by reports of officers misusing the cameras and ordered a state privacy review. That's the kind of misuse that cost Chandler, Arizona, its 40 cameras earlier this month, after an audit flagged an officer looking up a family member. The retention number will land wherever the legislatures land it. My point is narrower. Whatever the number is, a shorter window isn't a loss for the people doing the work. It's a push toward the work they signed up for.
I'd like to hear from people running these rooms today. How much of your daily value is back-search, and how much is live? If the archive went to seven days tomorrow, what would break, and what would you have to build instead? I have my own guesses. I'd rather have yours.