Perplexity has launched Brain, a memory system for its Computer agent, and the interesting part is what it chooses to remember. Where consumer AI memory has mostly meant recalling a user's preferences, the kind of thing that makes a chatbot remember your name or your writing style, Brain remembers the agent's own work: what it did, what went wrong, and what it learned. The company announced it on June 18 as a research preview.
At its center is a context graph, which Perplexity describes as an automatically built wiki that maps the people, projects, ideas, and moving parts in a user's world. As the agent works, it logs tasks, results, and corrections, and every entry in the graph links back to the session, file, or document it came from. That traceability is the point: rather than a vague sense of what it knows, the agent can show where a given piece of knowledge originated.
The self-improving part happens overnight. Each night, Brain reviews the day's sessions and synthesizes pending corrections into reusable lessons, and on later days the agent retrieves the relevant lessons before it starts a similar task. In effect, the system is built to stop repeating its own mistakes, learning from being corrected the way a new hire would, instead of starting every task from a blank slate.
Perplexity reports first-party gains from the approach: a 25 percent increase in answer correctness on repeated tasks, a 16 percent improvement in recall, and a 13 percent reduction in cost on work that depends on history. For now Brain is a research preview available only to Perplexity Max and Enterprise Max subscribers, and there is no public Brain API. The company has shown illustrative pseudocode but has not shipped a way for outside developers to build on it yet.
The reason this is worth watching is that it points at where agent memory is going. Remembering a user is easy and mostly cosmetic, but remembering work, grounding each memory in its source, and consolidating lessons so the system improves over time is the harder and more useful problem, and the overnight synthesis step is a recognizable nod to how consolidation works in memory more broadly. The caveats are the standard ones for a preview: the numbers are Perplexity's own, there is no independent evaluation, and access is limited. But an agent that gets measurably better at a task the second time because it remembered what it got wrong the first time is a more meaningful kind of memory than one that just remembers your name.