Self-improving Memory for Agents

A popular mental model in AI is to think of an agent as a worker. A new capability might be equivalent to “a recent college graduate,” or “a second-year analyst.” This is not the best way to think about AI capabilities, but it does illustrate a critical concept of AI at work: whenever you hire any worker, you expect them to learn and develop on the job. 

AI agents should do the same.  This begs a closer look at what “memory” really means in AI.  

Introducing Brain

Today we are launching a wholly new approach to memory in AI. We’re calling it Brain. 

Brain is a self-improving memory system. It builds a context graph of the work Computer performs. At set intervals, such as overnight, Brain reviews the context graph and teaches itself how to do the work better. 

The more work you do, the better and more efficient Brain makes your Computer. 

A better model for AI memory

There are two axes of importance for memory in AI. The first is what the memory is about, and the second is what the memory is for. 

Traditionally, AI memory has been about you, the user. Your preferences, tastes, working styles, contacts, role, and more. Brain pioneers a much more effective model for AI agents: Brain remembers what the agent did. It remembers what worked and what failed, and what corrections got made to the work. It learns to do better work. 

Second, AI memory about the user serves the purpose of helping you feel more engaged with the AI agent. Work memory like Brain serves the purpose of helping the agent get better at the job. This is the most important purpose of memory. 

Together, this approach to memory improves Computer’s performance over time. Every day, Computer with Brain has a better starting point for each new task. That’s because Brain creates a fresh map of what the user is most likely hoping to accomplish based on past sessions, what’s worked, and what hasn’t. 

That means Computer can get to answers faster, access the most reliable sources, and avoid wasting time and tokens.

Recursive self-improvement

Brain gets better as you use Computer. Agents become more effective at updating context as they learn the projects, connectors, artifacts, and other sources that lead to the best outputs.

They also learn from their mistakes, remembering when a user has made a correction or when a source was a dead end. That results in fewer turns, fewer model calls, and better outputs.

This feedback loop is what makes Brain continuously self-improving. 

For users, the agents get better at understanding the context they need to know to do the work. It also means that the current token usage is an investment in more efficient token usage later.   

The context graph 

Brain forms a living context graph for Computer by creating a traceable graph that helps Computer understand the user’s world, learn from their work, and apply that work to future tasks.

The context layer takes the form of an LLM wiki that’s automatically loaded onto the agent sandbox. A user's LLM wiki pages reflect the ideas, people, projects, and other elements that make up their world, allowing Computer to traverse this web of personal information. 

The wiki is incrementally updated by the Brain system overnight as it synthesizes the user’s sessions along with connector results, changes in source documents, and corrections made. 

That persistently refreshing context gives Computer stronger signal on what to do, where to look, and how to deliver outputs for a user.  

Improved performance 

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%. And the results for people who use Brain longer improve as the agents learn their world. 

Brain also shows its work, like other outputs generated by Perplexity. Every memory entry links back to the session, file, or source that it came from.

Powering proactive AI

Only through continuous learning does Computer become the proactive AI that users and businesses ask us for. Agents learn and remember the work and how to get better at it.

The frontier AI is the individual system you build for AI in your own organization. When agents actively learn from the work we do, they can identify opportunities we didn’t ask about, or flag a problem before anyone has noticed.

This version of Brain is just the beginning. We will announce new capabilities soon.

Brain is rolling out today to Max and Enterprise Max subscribers in Research Preview.