Summary
- Perplexity has unveiled Brain, a memory framework for its Computer agent that constructs a context graph from previous sessions, sources, and corrections, synthesizing this information overnight into a personalized LLM wiki to enhance future tasks.
- Initial metrics from Perplexity indicate that Brain improves answer accuracy by 25% on repeated tasks, boosts recall by 16%, and reduces the cost of context-intensive tasks by 13%.
- Brain is being rolled out today in Research Preview for Max ($200/month) and Enterprise Max subscribers, with memory options available under "Customize" in the sidebar.
Today, Perplexity introduced Brain, a memory system designed for its Computer AI that learns from prior interactions. Rather than simply remembering user details, it focuses on logging the actions taken by the agent.
According to Perplexity, “With Brain, Computer begins each task equipped with the complete context of your projects, decisions, and sources instead of starting from scratch. Each memory is linked to the session, file, or source it originated from, ensuring full transparency and control.”
Every time the Computer completes a task, Brain incorporates it into a context graph. This graph monitors which connectors were utilized, which sources were reliable, the corrections made by the user, and what did not yield results. At regular intervals—typically overnight—Brain synthesizes the graph and updates a personalized LLM wiki that is loaded into Computer's workspace before the next task. Each memory entry is connected to its source session or file, allowing users to trace decisions back to their origin.
The concept is simple: while most AI memory focuses on the user—tracking preferences and habits—Brain centers on the work itself. It records what the agent attempted, what was corrected, and which sources were beneficial, making it a more practical type of memory for systems designed to accomplish tasks.
Perplexity's early metrics suggest that Brain enhances answer correctness by 25% on tasks previously handled by Computer, improves recall by 16%, and lowers the expense of context-dependent tasks by 13%. These figures are derived from internal assessments rather than independent benchmarks. However, the trend is logical: an agent that starts each day aware of which sources were ineffective the previous week will spend fewer resources rediscovering that information.
Source: Perplexity AIFamiliar Ground
While some may find this advancement intriguing, others might wonder if it truly represents a breakthrough. In essence, Perplexity is adapting a specialized application for a wider audience.
OpenClaw has been implementing similar concepts for months, accumulating over 379,000 stars on GitHub. It employs markdown files and a SQLite database with FTS5 full-text search to maintain context across sessions. With the Mem0 plugin, memory capture occurs automatically at the system level, surviving restarts and context compaction.
In April 2026, OpenClaw introduced "providence labels," categorizing each stored memory as observed, user-confirmed, model-inferred, or imported from a transcript, allowing the agent to assess the reliability of any fact.
Hermes, a self-improving agent from Nous Research that launched in February 2026, takes this further. After completing a task, Hermes assesses the results, extracts reusable reasoning patterns, and documents them as skill files in plain markdown. The next time it encounters a similar issue, it can utilize the skill rather than starting from scratch.
Hermes also incorporates skills with similar functions (like Obsidian Mind) aimed at enhancing the agent's personal utility.
Both of these tools are self-hosted, allowing users to run them on their own hardware, ensuring data privacy and control. This contrasts with Brain, which operates entirely within Perplexity's ecosystem and is backed by a multibillion-dollar company.
Target Audience and Limitations
Brain is designed for users already subscribing to Perplexity Computer at $200 per month. For those engaged in recurrent tasks—such as competitive analysis, weekly reporting, or research referencing previous work—the upgrade could be significant, as the agent avoids redundant efforts in each session.
However, it is important to note that Brain does not function as a local memory tool under user control. The context graph, LLM wiki, and all session histories are maintained by Perplexity's infrastructure. Users can see what is stored, but they do not own it. Those seeking full data sovereignty might find better options with Hermes or OpenClaw, which offer plugins and skills like Mem0, Honcho, Obsidian Mind, or Hindsight, keeping data on personal hardware.
It is also crucial to clarify what "self-improving" means in this context. Brain enhances Computer's performance on tasks it has already completed for the user but does not make the underlying models more intelligent. The challenge of cross-domain generalization—applying knowledge gained from financial research to a coding task—remains unresolved and is not addressed by Brain.
Brain is available in Research Preview as of today for Max and Enterprise Max subscribers, with new features expected soon, though no timeline has been provided.