AI Agent · Perplexity
Perplexity Brain — agent that learns overnight
Agents used to start from scratch every session, often stepping on the same mistake the next day. Captures the Computer agent's own work history (successes, failures, recoveries) as a context graph and reviews it asynchronously each night — that's Perplexity's new Brain.
The zero-start limitation
AI agents couldn't carry memory across sessions — a hard constraint. A procedure you debugged on Monday was forgotten by Tuesday, and the same investigation started from scratch.
In repeated workflow automation, this amnesia was the biggest friction. On complex multi-step tasks, watching the agent retrace the same wrong turns and then manually correcting it became a recurring cost.
How Brain works
Captures the Computer agent's own work history (successes, failures, recoveries) as a context graph and reviews it asynchronously each night.
After the Perplexity Computer agent finishes a day's work, Brain analyzes that history. Which steps succeeded, where it failed, and how it recovered are stored as a directed graph. The next morning's session references this graph, letting the agent benefit from prior experience.
Unlike injecting memory into prompts, this functions as external memory that doesn't consume context window space.
The conditions for getting smarter
Recurring workflows get faster on the second run onward. One-off searches won't feel different.
People automating the same process daily — data collection, report generation, templated emails via the Computer agent — will notice Brain after a few runs. One or two initial sessions build up learning data, and from then on the agent steers around known failure patterns.
For users whose primary use is search and single Q&A, the context graph that accumulates is thin and no perceptible improvement appears. Brain's value scales directly with repetition frequency.
Source: Perplexity official