Perplexity's Brain Gives Its AI Agent a Memory

Perplexity launched Brain on June 18, a memory system that lets its Computer agent learn from its own work instead of starting cold every time. The twist is what it chooses to remember: not facts about you, but what the agent did — which sources paid off, which led nowhere, and which corrections you made — so it can reuse that the next time you ask for something similar.

Under the hood it builds what Perplexity calls a context graph, described as an LLM wiki: pages for the people, projects, and topics in your world that load automatically into the agent's sandbox. At set intervals — overnight, for example — Brain reviews recent sessions, connector pulls, and document changes, then folds the lessons back in. The stated goal is blunt: stop relearning the same context twice.

Why this matters

The numbers Perplexity reports are first-party and early, so read them as a direction rather than a guarantee: +25% answer correctness on repeated tasks, +16% recall, and −13% cost on work that needs historical context. The company says the gains compound the longer you use it. That is the real shift. Agents have mostly behaved like goldfish, forgetting everything between sessions; a working memory is what turns a clever demo into something you can lean on for recurring work — the weekly report, the same set of clients, the research you redo every month.

For now Brain is in Research Preview for Perplexity Max and Enterprise Max subscribers, inside the Computer agent. If you don't pay for those tiers you won't see it yet — but the idea will spread, because every agent vendor is chasing the same problem: an assistant that forgets everything by morning is hard to trust with anything that takes more than one sitting.

I've lost count of how many agent sessions I've spent re-explaining my own project from scratch. A memory that survives the night is the unglamorous feature I actually want — far more than another point on a benchmark.