--- title: "Perplexity's Brain Cuts Agent Task Costs by Teaching Itself Overnight" subtitle: "Perplexity's Brain gives its Computer agent a self-improving memory that gets cheaper and smarter with every task you run" url: "https://alphasignal.ai/news/perplexity-s-brain-cuts-agent-task-costs-by-teaching-itself-overnight" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Perplexity" company_url: "https://alphasignal.ai/company/perplexity_ai" published: "2026-06-18T16:14:35.000Z" updated: "2026-06-18T16:42:32.723Z" type: "News" topics: - "Retrieval" - "Development" subtopics: - "Rag" - "Document Understanding" upvotes: 3114 read_time: "6 min" source_url: "https://www.perplexity.ai/hub/blog/self-improving-memory-for-agents" comments: 0 --- # Perplexity's Brain Cuts Agent Task Costs by Teaching Itself Overnight > Perplexity's Brain gives its Computer agent a self-improving memory that gets cheaper and smarter with every task you run Every time you start a new session with an AI agent, you pay a hidden tax: re-explaining your project, your preferences, your sources, and your past decisions. [Perplexity's Brain](https://www.perplexity.ai/hub/blog/self-improving-memory-for-agents) is designed to eliminate that cost entirely, and then some. Brain is a continuously learning memory system built directly into [Perplexity Computer](https://www.perplexity.ai/products/computer), the company's multi-model agent platform. The key claim: it doesn't just remember you, it remembers the **work**, and gets better at doing it over time. ## The problem with how agents forget Most AI memory systems are built around the user. They store your preferences, your name, your role. Brain pioneers a different model: it remembers what the agent did, what worked and what failed, and what corrections got made to the work. That's a meaningful shift. User-preference memory helps you feel recognized. Work memory helps the agent actually improve. The practical consequence of stateless agents is expensive. Without persistent work context, Computer starts every task from scratch, re-discovering the same sources, re-making the same decisions, and burning tokens on context that already existed. Brain is Perplexity's answer to that. ## How Brain actually works Brain is built around what Perplexity calls a **context graph** , a structured, traceable map of the user's work history. The context layer takes the form of an LLM wiki that's automatically loaded onto the agent sandbox. A user's LLM wiki pages reflect the ideas, people, projects, and other elements that make up their world, allowing Computer to traverse this web of personal information. An LLM wiki, in this context, is a set of automatically maintained pages , think of it as a living internal knowledge base that the agent writes and updates itself, not one you have to curate manually. The wiki is incrementally updated by the Brain system overnight as it synthesizes the user's sessions along with connector results, changes in source documents, and corrections made. The overnight refresh is deliberate. Rather than updating memory in real-time (which adds latency and noise), Brain reviews the context graph at set intervals and teaches itself how to do the work better. It's a batch learning loop, not a streaming one. What makes it self-improving rather than just persistent: - Agents learn from mistakes, remembering when a user has made a correction or when a source was a dead end. - That results in fewer turns, fewer model calls, and better outputs over time. - Every memory entry links back to the session, file, or source it came from , giving full auditability of why the agent believes what it believes. - The more you use it, the better the starting context for each new task , the current token spend is an investment in cheaper future runs. ## The numbers behind the claim Early measurement results show that Brain increases answer correctness by 25% on tasks Computer has seen before, and recall goes up by 16%. Those are gains on tasks with historical context , not general benchmarks. The same results show Brain cuts the cost of tasks that require historical context by 13%. That last number is the one worth paying attention to: a memory system that actively reduces token spend as it matures is a different value proposition than one that just adds context overhead. For context on where Perplexity's memory architecture has been heading: the enhanced memory engine was able to recall important information in 95% of cases, up from 77% previously, while making half as many memories. The philosophy across both updates is the same , fewer, higher-quality memories beat noisy, exhaustive ones. ## The context graph vs. RAG Most agent memory systems today are built on **RAG** (Retrieval-Augmented Generation) , a technique where relevant documents are fetched from a vector database and stuffed into the prompt at query time. RAG is good at finding relevant chunks of text, but it doesn't track how facts relate to each other, and it can't learn from corrections. An LLM memory graph represents knowledge as a network of interconnected entities and their relationships, allowing for more intuitive querying and deeper reasoning, enabling AI agents to recall and synthesize information contextually. The principle Brain is built on is that memory shouldn't be retrieval , memory should be synthesis. Instead of fetching documents, Brain builds a structured understanding of what the agent has done and what that means for future tasks. The overnight synthesis step is where that distinction becomes concrete. ## What it's good for Brain is most valuable in workflows that repeat or compound over time. Practical use cases where it pays off: - **Long-running research projects** , Brain remembers which sources were reliable, which were dead ends, and what conclusions were already drawn. - **Recurring business workflows** , weekly reports, competitive monitoring, client research , where the same context applies session after session. - **Multi-step coding or deployment tasks** , where past decisions about architecture, naming, and tooling should carry forward automatically. - **Anything with corrections** , if you've told Computer to stop using a certain source or to prefer a certain format, Brain remembers that without you having to repeat it. It's less useful for one-off tasks with no prior history, where there's simply nothing for Brain to have learned yet. The value compounds with usage. ## The bigger picture Perplexity is no longer just an AI search engine , the company is pivoting from search to full agentic AI. Computer is a unified workspace that orchestrates roughly 19 different AI models to plan, delegate, and complete complex workflows end to end. Brain is the memory layer that makes that system stateful. The broader industry assumption Brain challenges is that agent memory is a personalization feature. Perplexity is betting it's an **operational efficiency feature** , one that should reduce cost and improve accuracy on every repeated task. If that framing holds, the question for every agentic platform becomes: what is your agent learning between sessions, and how do you measure it? ## Access and pricing Brain is rolling out today to Max and Enterprise Max subscribers in Research Preview. Perplexity Computer costs $200 per month or $2,000 per year as part of the Perplexity Max subscription plan. Every Max subscriber receives 10,000 credits per month, and each task consumes credits based on its complexity , a simple task might cost around 30 credits, while an extended coding session can burn through thousands. To access Brain, you can find it and your saved memories under "Customize" in the sidebar. Since it's a research preview, expect rough edges , but also expect the performance gains to be most visible on workflows you've already been running through Computer for a while. ## Sources - [X post by @perplexity_ai](https://x.com/perplexity_ai/status/2067642139014742348) - [Self-improving Memory for Agents — Perplexity Blog](https://www.perplexity.ai/hub/blog/self-improving-memory-for-agents) - [Perplexity Ships Computer and Memory Upgrades — The Search Signal](https://thesearchsignal.com/perplexity-computer-memory-upgrades/) - [Perplexity What We Shipped — February 6th, 2026](https://www.perplexity.ai/changelog/what-we-shipped---february-6th-2026) - [LLM Memory Graph: Connecting Knowledge for Smarter AI Agents](https://aiagentmemory.org/articles/llm-memory-graph/) - [Perplexity Computer Cost: Pricing, Credits & Plans (2026)](https://www.sentisight.ai/how-much-perplexity-computer-cost/)