Perplexity adds Brain memory to Computer agent

Perplexity launched Brain today, a memory system for its Computer agent that records past sessions into a context graph and synthesizes those records into a compact personal LLM wiki. The wiki is loaded into the agent’s sandbox before each new task. The feature is available in Research Preview for Max and Enterprise Max subscribers.

When Computer completes a task, Brain records which connectors and sources were used, what corrections the user made and which attempts failed. Those entries become nodes in a context graph. At scheduled intervals, overnight by default, Brain synthesizes the graph into a concise LLM wiki that Computer consults at the start of the next session.

Perplexity reports early internal metrics showing that Brain improves answer correctness by 25% on repeated tasks, raises recall by 16% and reduces the cost of context-heavy tasks by 13%. The company notes these figures are internal measurements rather than independent benchmarks. Perplexity also says each stored memory links back to the original session, file or source and that users can view memories in the Computer sidebar under “Customize.”

Perplexity positions Brain for users who run recurring, context-rich work such as competitive monitoring, weekly reporting and repeat research tasks. The system is designed so the agent does not need to rediscover prior failures or prior successful approaches each session.

Brain differs from self-hosted projects that pursue similar persistence. Tools such as OpenClaw and Hermes persist context on users’ hardware using local files, databases and plugins. Those projects can include automated memory capture, provenance labels and extracted skill files that load prior solutions. Brain runs within Perplexity’s infrastructure and stores the context graph, wiki and session history in the company’s ecosystem rather than on customer hardware.

Perplexity provides controls and visibility into what Brain stores but does not transfer ownership of the underlying storage. Customers with strict data sovereignty or on-premises requirements may prefer self-hosted alternatives that keep data local.

Perplexity also notes limits to Brain’s scope: the system helps Computer perform better on tasks it has already handled for a user, but it does not change the underlying language models or guarantee that learning in one domain will generalize to unrelated domains. Perplexity says additional capabilities are planned, without a public timeline.