Perplexity Launches AI Agent with Learning Capability

Perplexity has launched a new memory system called Brain for its AI agent, Computer, which aims to enhance the agent's efficiency and accuracy by learning from previous tasks. This feature builds a context graph during interactions, documenting what worked, what needed corrections, and linking back to the relevant sessions and sources.

Brain adds entries to a context graph after each task, which it synthesizes overnight to create an updated personal large language model (LLM) wiki that prepares Computer for upcoming tasks. This function allows the AI to begin with comprehensive context rather than starting from scratch. According to Perplexity, this approach provides users with full transparency and control over each memory entry, which connects to the session or file from which it originated.

Perplexity claims that early internal metrics reveal that Brain boosts correctness of answers by 25% on repeated tasks. It also improves recall by 16% and decreases the cost of context-heavy tasks by 13%. Although these numbers are based on their internal analysis rather than independent tests, they suggest a significant potential for increasing efficiency and reducing the resources required for tasks.

Adopting a memory system focused on work outcomes rather than user preferences could be a shift in how AI agents approach task management. Traditional AI memory often emphasized personal user data, while Brain concentrates on actionable memories that highlight what strategies succeeded or failed. This innovative mindset could pave the way for more effective AI applications across various sectors.

As Brain rolls out in a Research Preview for Max and Enterprise Max subscribers, users can access their memories through the "Customize" option in the sidebar. The feature's implementation and real-world performance will be noteworthy to observe as more users begin to utilize it.