Perplexity is shifting the focus of AI memory from user preference to agent performance with the launch of “Brain,” a self-improving system designed to enhance how its Computer tool performs tasks. Unlike traditional AI capability models that might equate progress to the skillset of a recent college graduate, Brain builds a context graph of the work Computer performs, allowing the AI to learn at set intervals and become more efficient. The system doesn’t prioritize remembering a user’s tastes or contacts; instead, it focuses on what the AI itself did, remembering what worked and what failed. “Work memory like Brain helps the agent get better at the job,” Perplexity explains, and early measurements show Brain increases answer correctness by 25 percent on tasks Computer has seen before while reducing costs by 13 percent.
Brain: Self-Improving Context Graph for AI Agents
This approach fundamentally shifts the purpose of AI memory, moving away from maximizing user engagement toward enhancing the agent’s efficiency. The core innovation lies in Brain’s ability to build a self-improving memory system, constructing a context graph of the work Computer performs; the more work done, the more refined and efficient Brain becomes. Perplexity emphasizes two key aspects of memory: what it contains and its intended purpose, noting that traditional AI memory centers on the user, while Brain remembers what the agent did, specifically what succeeded and failed, and what corrections were made. Early measurements demonstrate a 25 percent increase in answer correctness on previously encountered tasks, alongside a 16 percent increase in recall, and a 13 percent reduction in costs for tasks requiring historical context; these gains are amplified with extended use as the agent learns the user’s specific world. Brain creates a traceable graph, effectively an LLM wiki, that’s automatically updated overnight with session data, connector results, and user corrections, providing Computer with a stronger signal for future tasks and ultimately enabling proactive AI capable of identifying opportunities and flagging potential problems.
Early measurement results show that Brain increases answer correctness by 25% on tasks Computer has seen before. And recall goes up by 16%. The same results show Brain cuts the cost of tasks that require historical context by 13%.
AI Memory Shifts Focus to Agent Performance
The prevailing paradigm of artificial intelligence capability assessment, often framed around benchmarks equivalent to a recent college graduate, is undergoing a fundamental shift as developers prioritize self-improvement over mimicking human skill levels. Perplexity is developing this change with the launch of “Brain,” a novel memory system designed not to catalog user preferences, but to meticulously record and analyze the performance of the AI agent itself. This system constructs a context graph of the work Computer performs, enabling the agent to learn from both successes and failures, and refine its approach over time. The system’s architecture relies on an “LLM wiki” automatically loaded into an agent sandbox, forming a living context graph that allows the AI to understand a user’s world and apply past learnings to future tasks.
This continuous learning loop not only enhances efficiency, reducing the cost of context-dependent tasks by 13 percent, but also positions the AI to proactively identify opportunities and potential problems, moving beyond reactive task completion towards a truly intelligent assistant. Brain is currently available in Research Preview to Max and Enterprise Max subscribers, with further capabilities planned for future release.
Brain Improves Computer’s Accuracy, Recall, and Cost Efficiency
Perplexity is fundamentally reshaping how AI agents retain information, moving beyond user-centric memory to focus on an agent’s own performance history. Unlike conventional AI capability assessments that benchmark progress against the skillset of a recent college graduate, the company is prioritizing a system of continuous self-improvement through what it calls Brain. Brain isn’t designed to catalogue a user’s preferences; instead, it concentrates on remembering specifically what strategies proved effective and which required correction. The core principle driving Brain is to enhance the agent’s ability to perform tasks efficiently, rather than simply maximizing user engagement. Recall also increases by 16 percent, indicating a stronger ability to retrieve relevant information when needed. Perplexity highlights that these gains aren’t merely immediate boosts, but rather represent an investment in future efficiency, as Brain learns to optimize token usage over time. The company also reports a 13 percent reduction in the cost of tasks requiring historical context, suggesting that Brain’s self-improvement loop translates directly into economic benefits.
