Perplexity equips its agent Computer with Brain, a self-improving memory.

Perplexity launches Brain, a memory system for its Computer agent that cuts costs by 13% and boosts accuracy by 25% via a self-improving context graph.

Perplexity is adding a memory system called Brain to its Computer agent, which is designed to self-improve with use. Whereas an AI's memory typically focuses on the user—their preferences, contacts, and style—Brain concentrates on the agent's work: what it attempted, what succeeded or failed, and the corrections it received. The objective is no longer to personalize the relationship but to help the agent work more effectively.

In practice, Brain builds a context graph of Computer's work, structured as an internal wiki loaded into the agent's sandbox and incrementally re-synthesized, often overnight, from sessions, connectors, document modifications, and corrections. Each new task then begins with the context of projects and sources, rather than from scratch.

The loop is designed to be recursive: the agent identifies which sources and projects lead to the best results, learns from its dead ends, and consequently reduces back-and-forth interactions and calls to the model, thereby lowering token expenditure. Each memory remains traceable back to the session, file, or source from which it originated. Based on its initial tests, Perplexity reports a 25% increase in accuracy and a 16% improvement in recall on previously encountered tasks, with a 13% cost reduction when historical context is required. Brain is being released as a Research Preview for Max and Enterprise Max subscribers.