Perplexity introduces Brain, a self-improving memory for its AI agent

Perplexity has launched a new memory system called Brain for its Computer agent — a way for the AI to learn from its own actions and improve over time.

Unlike typical AI memory that stores user preferences and personal details, Brain focuses on the agent's own behavior and task performance. Each time the agent runs a task, Brain automatically builds a context graph in the background, tracking which tools were used, which sources succeeded or failed, the outcomes, and any user corrections. That context graph becomes a dynamic AI knowledge base, loaded into the agent's sandbox environment the next time it starts.

To avoid getting stuck on a single task, Brain performs incremental synthesis overnight — merging all the day's sessions, connected app outputs, and feedback, then pruning and optimizing the knowledge base.

Preloading this history helps the agent avoid repeating mistakes, flag potential issues, and cut down on redundant attempts. Perplexity shared test data showing that Brain improved answer accuracy by 25% in repeated tasks, boosted information recall by 16%, and reduced token consumption by 13% in tasks relying on historical context.

The performance gains should compound the longer the system is used. Every memory in Brain is traceable back to its original session or file for transparency. The new system is now available as a research preview for Perplexity Max and Enterprise Max subscribers.