Perplexity Launches Brain: Context-Graph Memory Cuts Computer Agent Cost 13%, Boosts Accuracy 25%
Perplexity launched Brain on June 18, a memory system for its Computer agent built on a different premise than every other AI memory product in the market. Most memory systems remember the user. Brain remembers the work.
Every completed Computer task feeds into a context graph tracking which connectors delivered results, which sources failed, what corrections the user made, and what approaches stalled. At set intervals, overnight by default, Brain synthesizes that graph into a personal LLM wiki, then loads it automatically into the agent’s sandbox before each new task starts.
The practical consequence: Computer does not restart from scratch every session. It knows which Notion pages contain the relevant policies, which web sources misled it last time, and which SQL connector reliably answered questions about Q2 revenue.
What the Context Graph Tracks
The memory structure is a graph, not a flat list. It preserves relationships between elements of a user’s work: a decision made in one session can connect to a file from a connector and a result from a prior search. Sessions, corrections, and connector outputs all feed in. The wiki self-updates overnight, so by the time the next task starts, Brain’s context layer reflects everything that changed since the last refresh.
Every memory entry is traceable back to its origin session or file. Users can inspect what Brain knows and where it came from, but the data lives on Perplexity’s infrastructure. There is no export or self-hosting option.
Numbers
Perplexity’s internal benchmarks on the feature before external release:
| Metric | Improvement |
|---|---|
| Answer correctness (repeated tasks) | +25% |
| Recall | +16% |
| Cost per task (context-heavy work) | -13% |
These are first-party figures. No independent benchmark exists yet.
The gains are structurally sound. An agent that opens each task knowing which sources it spent two sessions validating last week will waste fewer model calls re-establishing that baseline. The 13% cost reduction reflects token savings from better context injection at task start rather than cheaper underlying models.
Availability
Brain is in Research Preview for Perplexity Max ($200/month) and Enterprise Max subscribers. It surfaces under Customize in the sidebar. Perplexity gave no timeline for broader rollout.
The stated target is recurring, continuous work: competitive monitoring, weekly reporting, research tasks that reference prior runs. For one-off queries, it adds nothing.
The Data Trade-Off
The context graph, LLM wiki, and all session history sit on Perplexity’s infrastructure. Users can read what is stored but cannot take it elsewhere. Teams with data sovereignty requirements or compliance constraints are not the target audience.
Where This Fits
OpenAI’s Dreaming V3 synthesizes ChatGPT memories from conversation history, reaching 82.8% recall on its own benchmarks. Anthropic’s Claude Managed Agents added persistent memory in public beta this month, with Rakuten reporting 97% fewer errors. Both approaches centre on user context, preferences, and conversation history.
Brain’s distinction is the frame of reference. It organises memory around task outcomes, connector results, and agent decisions rather than who the user is. No competitor has shipped a context graph structured that way at this scope. Whether that distinction drives retention among the $200/month subscriber base is the open question.