Perplexity’s Brain learns from agent work to improve itself overnight
Perplexity is shifting how AI memory works. Instead of storing user preferences to boost engagement, its new Brain system records what the agent actually does, then retrains itself overnight to do that work better next time. The feature, launching today for Max and Enterprise Max users, treats memory as a performance tool rather than a personalization engine.
From personal profiles to performance logs
Most AI assistants build memory around the user: tastes, contacts, roles. Brain turns that axis sideways. It keeps a traceable context graph of the agent’s own actions—sessions, connectors, document changes, corrections made. Each memory entry links back to its source, so users can audit how the system reached its conclusions. The result is a living LLM “wiki” that Computer consults before acting, updated automatically while you sleep.
A feedback loop that cuts cost and improves quality
Every correction you make, every dead-end source you reject, feeds into Brain’s nightly retraining. The system learns which projects, connectors, and artifacts lead to the best outputs and which steps waste tokens. In Perplexity’s early tests, the changes translated into measurable gains: 25% higher answer correctness on familiar tasks, 16% better recall under similar conditions, and 13% lower token usage when historical context mattered. Performance continues to climb the longer a user keeps Brain active.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.