How the context graph works

Every task Computer performs plugs into a persistent context graph — a structured map spanning projects, connectors, files, artifacts, decisions, and sources . This graph grows with each session, creating a rich operational history that future tasks can draw from automatically.

Unlike simple conversation memory (which stores chat logs), Brain captures the agent's procedural knowledge: which workflows produce correct results, which approaches were corrected by the user, and which source patterns are most reliable .

The overnight self-improvement loop

Brain's most distinctive feature is its ability to improve autonomously. At set intervals — typically overnight — Brain reviews the entire context graph and updates its own understanding of how to do similar work better . This means:

• Learning from success: Patterns that led to correct outputs are reinforced.
• Learning from failure: Mistakes and user corrections are analyzed and used to adjust future behavior.
• Recursive improvement: As agents become more effective at updating the context graph, Brain's knowledge quality compounds .

The result is a system that gets smarter the more it's used. The gains are not just theoretical — Perplexity published early internal measurements directly on its blog :

These gains compound over time — users who use Brain longer see larger improvements as the agents learn their specific work patterns .

Why this is a shift from stateless to stateful agents

Brain represents a fundamental architectural shift in how AI agents operate. Most agents today are stateless — they start every task from scratch, with no memory of past work. Users must re-explain context, re-establish preferences, and re-train the agent on their workflows each time.

Brain makes agents stateful. Every new task plugs into the persistent context graph of prior projects, decisions, and corrections . Computer begins each task with full context instead of from nothing, eliminating redundant work and reducing errors from missing historical information .

This is distinct from simple conversation memory (which remembers chat history). Brain remembers the agent's operational knowledge — what procedures, tools, and source patterns produce correct results .

Which subscription tiers offer Brain

Brain is available as a research preview exclusively for Perplexity Max subscribers ($200/month) . It is rolling out first to Max, Perplexity's top consumer tier, which also offers unlimited Pro searches, full model lineup access (including GPT-5, Claude 4, Gemini), and early access to new features .

The feature is not available on Free, Pro ($20/month), or Enterprise Pro tiers at launch.

Where Brain fits in the Computer ecosystem

• Computer (launched February 2026) is Perplexity's "general-purpose digital worker" — an agentic platform that takes high-level goals in natural language, breaks them into subtasks, and executes them autonomously across the web, APIs, and the user's software stack . It orchestrates roughly 19 different AI models to complete end-to-end workflows .
• Brain is the memory layer for Computer. It makes Computer a learning system that gets more efficient and accurate the more it's used .
• Perplexity also has a separate personalization memory feature (rolled out earlier to Pro and Max users) that remembers user preferences and conversation history for the search/productivity assistant — Brain is about agentic work memory, not user preference memory .

How Brain compares to competing memory systems

Brain's differentiation is clear: it remembers the agent's work, not the user's chat — and it improves itself autonomously overnight .

• Anthropic (Claude): Has limited memory features for projects and custom instructions but no self-improving agent memory system. Brain's overnight self-review loop is unique.
• OpenAI (ChatGPT/Operator): Offers memory of user preferences and conversation history across chat, but OpenAI's agent (Operator) does not have a recursive self-improvement memory system that operates offline.
• Google (Gemini): Memory is primarily personalization-focused, not agentic work memory.
• Microsoft Copilot: Enterprise memory tied to Microsoft Graph (emails, documents), but not a self-improving agent memory that learns from task execution patterns.

Brain makes Perplexity a clear first-mover in stateful, self-improving agent memory — a category that most competitors have not yet addressed at the product level .

How the context graph works

Every task Computer performs plugs into a persistent context graph — a structured map spanning projects, connectors, files, artifacts, decisions, and sources . This graph grows with each session, creating a rich operational history that future tasks can draw from automatically.

Unlike simple conversation memory (which stores chat logs), Brain captures the agent's procedural knowledge: which workflows produce correct results, which approaches were corrected by the user, and which source patterns are most reliable .

The overnight self-improvement loop

Brain's most distinctive feature is its ability to improve autonomously. At set intervals — typically overnight — Brain reviews the entire context graph and updates its own understanding of how to do similar work better . This means:

• Learning from success: Patterns that led to correct outputs are reinforced.
• Learning from failure: Mistakes and user corrections are analyzed and used to adjust future behavior.
• Recursive improvement: As agents become more effective at updating the context graph, Brain's knowledge quality compounds .

The result is a system that gets smarter the more it's used. The gains are not just theoretical — Perplexity published early internal measurements directly on its blog :

These gains compound over time — users who use Brain longer see larger improvements as the agents learn their specific work patterns .

Why this is a shift from stateless to stateful agents

Brain represents a fundamental architectural shift in how AI agents operate. Most agents today are stateless — they start every task from scratch, with no memory of past work. Users must re-explain context, re-establish preferences, and re-train the agent on their workflows each time.

Brain makes agents stateful. Every new task plugs into the persistent context graph of prior projects, decisions, and corrections . Computer begins each task with full context instead of from nothing, eliminating redundant work and reducing errors from missing historical information .

This is distinct from simple conversation memory (which remembers chat history). Brain remembers the agent's operational knowledge — what procedures, tools, and source patterns produce correct results .

Which subscription tiers offer Brain

Brain is available as a research preview exclusively for Perplexity Max subscribers ($200/month) . It is rolling out first to Max, Perplexity's top consumer tier, which also offers unlimited Pro searches, full model lineup access (including GPT-5, Claude 4, Gemini), and early access to new features .

The feature is not available on Free, Pro ($20/month), or Enterprise Pro tiers at launch.

Where Brain fits in the Computer ecosystem

• Computer (launched February 2026) is Perplexity's "general-purpose digital worker" — an agentic platform that takes high-level goals in natural language, breaks them into subtasks, and executes them autonomously across the web, APIs, and the user's software stack . It orchestrates roughly 19 different AI models to complete end-to-end workflows .
• Brain is the memory layer for Computer. It makes Computer a learning system that gets more efficient and accurate the more it's used .
• Perplexity also has a separate personalization memory feature (rolled out earlier to Pro and Max users) that remembers user preferences and conversation history for the search/productivity assistant — Brain is about agentic work memory, not user preference memory .

How Brain compares to competing memory systems

Brain's differentiation is clear: it remembers the agent's work, not the user's chat — and it improves itself autonomously overnight .

• Anthropic (Claude): Has limited memory features for projects and custom instructions but no self-improving agent memory system. Brain's overnight self-review loop is unique.
• OpenAI (ChatGPT/Operator): Offers memory of user preferences and conversation history across chat, but OpenAI's agent (Operator) does not have a recursive self-improvement memory system that operates offline.
• Google (Gemini): Memory is primarily personalization-focused, not agentic work memory.
• Microsoft Copilot: Enterprise memory tied to Microsoft Graph (emails, documents), but not a self-improving agent memory that learns from task execution patterns.

Brain makes Perplexity a clear first-mover in stateful, self-improving agent memory — a category that most competitors have not yet addressed at the product level .