perplexity's agent now learns from its own work

Perplexity launched Brain today, a new kind of memory for its Computer agent. Most AI memory remembers you — your preferences, your style. Brain remembers the work: what the agent did, what worked, what failed, what you corrected. Overnight it reviews all of that and teaches itself to do the job better next time.

The payoff is measurable. Perplexity reports that on tasks the agent has seen before, Brain raises answer correctness by 25%, improves recall by 16%, and cuts the cost of context-heavy tasks by 13%. The longer you use it, the more it learns your world — and every memory entry links back to the session or source it came from, so you can see why it did what it did.

The catch, like a lot of frontier features right now: it's a research preview, limited to Perplexity's Max and Enterprise Max tiers — the $200-a-month end of the market, not where most independent operators here are. So this isn't a tool to go turn on today. It's a signal about direction: agent memory is moving from "remembers your preferences" to "remembers the work and compounds on it." An agent that gets measurably better at a recurring task every time you run it is a different thing than one that just answers — and it's the model the rest of the field will chase. If you're building with agents here, that shift is the part to watch.

This is a research preview on Perplexity's top-tier plans, not a broadly available tool. The point isn't the product — it's where agent memory is heading.