Perplexity Brain: AI Agents That Learn From Their Own Work

AI agent memory has long been focused on a single goal: knowing the user better. Your preferences, working style, usual contacts. What nobody had solved until now: remembering what the agent actually did, what worked, and how to operate better next time. Perplexity changed that with Brain, launched on June 18, 2026.

Direct answer: Brain is a self-improving memory system for AI agents. It builds a context graph of all work completed, synthesizes past sessions overnight to extract improvements, and loads this graph into the agent at the start of every new task. Results: 25% accuracy boost on known tasks, 16% recall improvement, and 13% lower cost on history-dependent tasks.

Two types of memory, two very different purposes

There are two ways to think about memory in an AI agent.

The first, and most common: user-centered memory. It retains your preferences, role, and usual contacts. Its purpose is to make the agent feel like it "knows" you. Useful for engagement, not for operational performance.

The second, the one Brain implements: work-centered memory. It retains what the agent did, what produced a good result, what failed, and what corrections the user made. Its purpose is for the agent to improve with every cycle.

This distinction matters. An agent that remembers your preferences feels more pleasant to use. An agent that remembers what worked yesterday delivers better answers tomorrow.

The context graph: a wiki the agent builds itself

Brain forms a living context graph: a traceable map of ideas, people, projects, sources, and artifacts that make up the user's working world.

This graph takes the form of an LLM wiki, automatically loaded into the agent sandbox at the start of each session. Every entry in the wiki traces back to the session, file, or source it came from. Full traceability is built in.

Overnight, Brain synthesizes the day's sessions: connector results, changes in source documents, corrections made by the user. It updates the wiki accordingly. The next day, the agent starts with a richer, more accurate graph tuned to what the user is trying to accomplish.

Practical result: the agent can retrieve the most reliable sources immediately, without searching, because Brain has already mapped the shortcuts from previous sessions.

Results published by Perplexity

Early measurements from Brain show:

• +25% answer accuracy on tasks the agent has seen before
• +16% recall rate (retrieving the right information)
• -13% cost on tasks that require historical context
• Continuous improvement: gains compound as the agent learns the user's world

The mechanism is self-reinforcing. Each user correction (a wrong source, an answer to rephrase) becomes training data. The agent remembers not just the correction but the context that made it necessary. Each session is therefore an investment in future efficiency.

What this means for production agent deployments

Brain represents a paradigm shift for production agents: from isolated sessions toward agents that accumulate operational knowledge.

For a company deploying a competitive intelligence agent or a document assistant, the question is no longer just "what is the performance on the first query?" but "what is the performance after 30 days of heavy use?"

An agent with persistent work memory can:

• Identify recurring gaps in the document base
• Automatically detect inconsistencies across sessions
• Reduce the number of LLM calls on repetitive tasks (hence the cost reduction)
• Surface opportunities or problems before being asked (proactive mode)

GDPR note for European teams: Brain is currently available to Perplexity Computer Max and Enterprise Max subscribers. The work data, corrections, and sessions that feed Brain are stored on Perplexity's infrastructure, hosted in the United States. For companies handling confidential or personal data, this means evaluating whether that data constitutes a cross-border transfer under Article 44 of the GDPR. Self-hosted solutions implementing similar logic (context graph and nightly synthesis on Elasticsearch or Qdrant, deployed on EU infrastructure) can reproduce this pattern with full data control and no dependency on a US provider.

TL;DR

Brain is the first mainstream implementation of agent memory centered on work done rather than user preferences. The LLM context graph, synthesized nightly, improves accuracy by 25% and reduces costs by 13% on history-dependent tasks. The pattern is reproducible in self-hosted form, which matters for European teams that cannot delegate their work data to US infrastructure.

Deploying an AI agent in production and want to implement persistent work memory on EU infrastructure? Let's talk.