Perplexity AI Agent Brain Learns From Mistakes, Boosting Accuracy and Cutting Context Costs

Perplexity has launched “Brain,” a memory system for its Computer AI agent, aiming to make the Perplexity AI Agent Brain smarter the more it is used. Instead of storing personal details, Perplexity AI Agent Brain logs what the agent actually did—what worked, what failed, and which sources and corrections were involved. Brain builds a context graph after each completed task, linking every memory entry back to the original session, file, or source for traceability and user control. At set intervals (overnight by default), Brain synthesizes this graph into a personal “LLM wiki” that loads into the Computer sandbox before the next task begins. Perplexity’s early internal metrics claim that the Perplexity AI Agent Brain improves answer correctness by 25% on repeated tasks, boosts recall by 16%, and reduces the cost of context-heavy tasks by 13%. The company stresses these are internal results, not third-party benchmarks. Rollout: Brain is now available in Research Preview for Max ($200/month) and Enterprise Max subscribers, with memories accessible under “Customize” in the sidebar. Perplexity frames Brain as bringing a niche AI-memory approach to a mainstream product. The article compares it with self-hosted agent memory efforts such as OpenClaw (using markdown + SQLite FTS5, plus Mem0-style plugins) and Hermes (skills written as markdown). It also notes a key limitation: Brain improves task performance for recurring workflows, but does not make the underlying models themselves smarter. For traders: while this is not a direct crypto protocol change, it signals accelerating adoption of “agent memory” features that can improve research workflows and potentially influence how market information is gathered and processed.