The Brain system changes AI memory from 'memorizing texts' to 'writing diaries'—recording how it works, where it messes up, how users criticize it, and automatically organizing and archiving at night. The 25% increase in accuracy actually comes from turning implicit experience into reusable explicit knowledge, and the intelligent agent is finally starting to learn from its mistakes.

CryptoWorld reports that Perplexity has launched a self-improving memory system for its intelligent agent Brain, which saves its own actions and task performance. During task execution, it automatically constructs a contextual graph and converts it into a dynamic knowledge base, loading the sandbox environment on the next startup. Incremental synthesis and pruning optimization are performed overnight. Tests show a 25% increase in accuracy, a 16% rise in information recall rate, and a 13% reduction in token consumption dependent on historical context. Currently, a research preview is open for Perplexity Max/Enterprise Max.