Perplexity Computer adds Brain context graph with +25% correctness on memory-heavy tasks

• Perplexity shipped Brain as a memory layer for Computer, and Perplexity's intro thread says it turns the agent into a stateful system built on a shared context graph instead of starting each run cold.
• According to Perplexity's benchmark post, Brain improved answer correctness by 25 percent, recall by 16 percent, and cost per task by 13 percent on work that depends on prior context.
• Arav Srinivas's rollout post says Brain pulls from prior sessions, connectors, and files, then refreshes itself overnight before feeding that context back into new Computer tasks.
• Memory entries stay inspectable: Perplexity's memory explainer says each one links back to the original session, file, or source, and Wes Roth's summary notes that users manage them from the Customize menu.
• Access is narrow for now. Both Perplexity's launch thread and Arav Srinivas's rollout post describe Brain as a research preview for Perplexity Max subscribers.

You can watch the launch demo via Perplexity's post, skim Arav Srinivas's rollout thread, and the company is already framing the feature less like chat history and more like a reusable context layer for agent runs Arav Srinivas on context graphs.

The launch pitch is simple: Brain is a continuously learning memory system inside Computer. Perplexity says every task plugs into the same context graph, so the agent accumulates project state, prior decisions, files, and source material over time instead of rebuilding context from scratch Perplexity's intro thread.

Srinivas adds one operational detail that matters: Brain is not only passively storing history. It updates itself overnight with fresh context, then feeds that refreshed graph back into subsequent tasks Arav Srinivas's rollout post.

Perplexity and Srinivas describe Brain as a graph built from several input types:

• Previous tasks and sessions Wes Roth's feature summary
• Projects and decisions Perplexity's benchmark post
• Files and sources Perplexity's memory explainer
• Connectors, in Srinivas's wording Arav Srinivas's rollout post
• Conversations, according to Wes Roth's feature summary

That combination makes Brain look closer to an agent memory substrate than a saved-chat feature. The context graph is the product.

Perplexity only attached metrics to tasks that require historical context, but the deltas are large enough to stand out:

• Answer correctness: +25% Perplexity's benchmark post
• Recall: +16% Perplexity's benchmark post
• Cost per task: -13% Perplexity's benchmark post

Perplexity's explanation for the cost drop is retrieval. Instead of reprocessing whole prior sessions, Brain can pull the most relevant memories for the current run Wes Roth's feature summary.

Perplexity says each memory links back to the session, file, or source it came from, which gives users a provenance trail for anything Brain recalls Perplexity's memory explainer. The company also surfaced Brain under Customize in the sidebar, where saved memories can be accessed and managed Perplexity's memory explainer.

That transparency claim is doing real work here. A persistent agent memory layer is much easier to trust when the retrieval has visible citations back to its origin.

Perplexity launched Brain as a research preview for Max subscribers, not as a general Computer default Perplexity's intro thread. The restriction showed up immediately in early reaction posts, including one summary post and another community recap, because it gates the most interesting part of the update behind the top tier.

Srinivas also gave the clearest statement of what Perplexity thinks it is building beyond this rollout. He argued that context graphs can become the deployment layer for agentic harnesses inside companies, because knowledge is fragmented across internal tools and a self-organizing graph can capture that tacit context over time Arav Srinivas on context graphs.

That makes Brain feel like more than a memory add-on. Perplexity is pitching it as the organizing layer underneath future agent workflows.