Perplexity Brain is memory for the work, not the user

Perplexity has launched Brain, a memory layer for its Computer agent.

The useful part is not that the agent remembers your name.

It remembers the work.

Every completed task goes into a context graph: connectors used, sources that held up, corrections made, dead ends, files, sessions and decisions.

Then, on a schedule, Brain turns that graph into a working memory that loads before the next task.

That is a much more serious idea than “personalization”.

Most consumer AI memory is built to make the product feel more familiar. It remembers preferences, tone, names, habits and surface-level context.

Brain is aimed at performance.

What worked last time?

What failed?

Which source was useful?

Which assumption did the user correct?

Perplexity says early internal numbers show 25% better answer correctness on tasks Computer has already handled, 16% better recall and 13% lower cost on context-heavy tasks.

Treat those as vendor numbers, not gospel.

But the mechanism is right.

An agent that has to rediscover everything every morning is not an operator. It is a very fast intern with amnesia.

The real agent stack needs memory that survives the context window.

Not just “remember Jason likes short sentences”.

Remember that a source was weak. Remember that a workflow failed. Remember that the correct file lives in the vault, not the local draft folder. Remember the decision chain, with provenance.

That is why this signal matters.

Agent memory is moving from novelty to infrastructure.

Sources: Decrypt, MarkTechPost