Perplexity has introduced Brain, a new self-improving memory system designed to make its Computer agents more accurate, efficient and context-aware over time. The feature, launched on June 18, 2026, is rolling out in Research Preview to Max and Enterprise Max subscribers.
Brain builds what Perplexity describes as a “context graph” of the work performed by Computer, tracking what the agent did, what succeeded or failed, and what corrections were applied by the user. At set intervals, such as overnight, Brain reviews this graph and teaches itself how to perform similar work better, so the agent starts each new task from a more informed position.
Unlike traditional AI memory systems that focus on user profiles and preferences, Brain is explicitly oriented around “work memory.” The system is designed to help agents improve at the job itself rather than primarily deepening personalization or engagement with the user. Perplexity says this allows Computer to reach answers faster, access more reliable sources, and avoid unproductive paths that waste time and tokens.
At the core of Brain is a continuously updated context layer that functions like an LLM-powered wiki of the user’s world. This wiki captures ideas, people, projects and other elements from prior sessions, connector results, changes in source documents and user corrections, and is automatically loaded into the agent sandbox for future tasks. Perplexity positions this as a “living context graph” that gives agents stronger signals on what to do, where to look and how to deliver outputs.
Early measurements cited by the company indicate that Brain increases answer correctness by 25 percent on tasks Computer has seen before, while recall improves by 16 percent. The same internal results suggest that Brain cuts the cost of tasks requiring historical context by 13 percent, with performance improving further as users work with the system over longer periods.
Perplexity emphasizes a recursive feedback loop in which agents become more effective at updating context as they learn which projects, connectors and artifacts lead to the best results. By remembering mistakes and dead-end sources, Brain is intended to reduce the number of interaction turns and model calls needed to complete a task, turning current token usage into what the company describes as an investment in more efficient usage later.
The company also frames Brain as a step toward more proactive AI systems that can learn continuously from user work. Perplexity argues that, as agents internalize more of an organization’s workflows and information, they will be better positioned to identify opportunities or flag problems without explicit prompts. This initial version of Brain is described as “just the beginning,” with Perplexity planning to announce additional capabilities in future updates.
Read Full Article: EPFO to Launch Instant PF Withdrawals via UPI and ATMs by the end of June
