The Perplexity brain gives its AI agent memory

On June 18th, Perplexity introduced Brain, a new feature allowing their Computer agent to build a memory of previous experiences. Instead of treating each interaction as a blank slate, the Computer agent can leverage past successes (and failures), such as which connectors were fruitful and which ones weren’t, or how you corrected its information. This way, when you come back asking for something similar, the Computer agent can draw upon these memories to provide better assistance.

Behind the scenes, Perplexity creates what they call a “context graph,” which they describe as an “LLM Wiki” where pages are created for everyone and everything relevant to your life, loading them directly into the Computer agent’s sandbox. Periodically (e.g., nightly), the Computer agent examines past interactions, connector pulls, and document edits, synthesizing these experiences back into the system. The aim here is simple: don’t repeat yourself.

1Why this matters

These metrics are self-reported and preliminary, so take them with a grain of salt: +25% accuracy on repeatable tasks, +16% recall, and -13% cost on tasks requiring historical context. According to Perplexity, these benefits become even stronger over time. And that’s the key difference. Up until this point, agents have been like goldfish; they forget everything once the session ends. But with a short-term memory, they become viable tools for repetitive tasks (like weekly reports, recurring clients, monthly research).

Brain is currently available in Research Preview for Perplexity Max and Enterprise Max customers within the Computer agent. Non-paying users won’t have access to Brain just yet, but rest assured, others will follow suit. Every agent platform wants to solve this issue: if the agent forgets everything by morning, how can you rely on it for tasks that require multiple sittings?

I can’t tell you how many times I’ve repeated myself during an agent interaction, explaining my project again. Give me a short-term memory that lasts longer than a day, and I’ll take that over any benchmark score.