Perplexity launched a product called Computer earlier this year. It is an AI agent that takes a goal, breaks it into steps, and completes those steps in the background using a mix of around 19 different AI models, pulling from whatever data sources you have connected. Think of it as a digital worker that can pull reports, draft documents, search your company's systems, and deliver finished outputs, all without you running each step yourself.
Brain is the new memory layer sitting underneath Computer. The difference from other AI memory systems is in what it chooses to remember. Most AI tools today store facts about the user: your role, your preferences, how you like your emails written. Brain stores facts about the work itself: which sources gave good results, which approaches led nowhere, and what corrections you made when the agent got something wrong.
Each night, Brain reviews what Computer did during the day. It synthesizes the sessions, the files consulted, the corrections made, and the dead ends hit. It then updates a structured record of the agent's work history. The next time a similar task comes up, Computer starts with that context already loaded, rather than approaching it from scratch.
This matters for cost as much as quality. Every time an AI agent reruns context it already gathered, it spends money doing so. The processing power required to run AI tasks is not free, and agents that do not remember past work repeat that spending every session. Brain's overnight synthesis converts that repeated spend into a stored lesson. Perplexity frames the early token cost as an investment in lower costs on future runs, and its early figures show a 13% cost reduction on tasks that depend on past context.
The accuracy numbers Perplexity reports, a 25% improvement in correctness and 16% improvement in recall on familiar tasks, come from internal testing during a research preview. They are directional, not audited. Any business evaluating this should test it against their own workflows before drawing conclusions.
Every memory entry in Brain links back to the specific session, file, or source it came from. That traceability is not just a technical feature. For anyone managing a team that uses AI agents for real work, being able to trace why the agent made a decision, and correct it at the source, is what separates a tool you can trust from one you cannot.
Brain is currently rolling out to Max subscribers at $200 per month and Enterprise Max subscribers at $325 per seat per month. Computer connects to over 400 business applications including Slack, Salesforce, SharePoint, GitHub, Gmail, and Notion. Enterprise plans include data isolation, SOC 2 Type II compliance, and spend controls.
The broader context here is worth noting. Perplexity grew its user base 3.7 times and revenue 4.7 times across 2025, and 92% of Fortune 500 companies reportedly have some usage of the platform, though much of that is individual employees using personal accounts rather than formal enterprise deployments. Brain is the company's clearest signal yet that it is trying to move from search tool to something closer to a standing digital worker that compounds in value the more you use it.
The question worth asking for any operator: how much of your team's AI usage today involves re-explaining the same context, re-researching the same ground, or correcting the same mistakes? If the answer is a lot, a memory system that stores and applies those corrections is genuinely worth watching.