Perplexity Rebuilds Projects With Brain Memory to Beat Stateless AI Agents

Perplexity has rebranded Spaces as Projects and rebuilt them into a persistent workspace inside Computer, the company's agentic AI platform. The rename is the least interesting part. What's actually new is a shared file system and deep integration with Brain, Perplexity's self-improving memory system, turning Projects into a continuously learning team workspace rather than a simple chat organizer.

Why stateless agents break down

Every AI agent forgets. Start a new session and you're back to square one, re-explaining context, re-uploading files, re-specifying preferences. For one-off tasks, that's tolerable. For ongoing work spanning days, teams, and dozens of files, it's a real bottleneck. Projects is designed to close that gap.

Each Project gives humans and agents a shared, persistent file system and a Brain that learns from completed work, so the next task starts with full context from the last.

A file system that acts like a teammate

Computer works in the file system the way a collaborator would. It opens files, reads their latest versions, makes edits, and saves changes back to the same files rather than producing a new artifact you have to manually reconcile with the previous one. That's a meaningful shift from the typical agent pattern.

The file system is hierarchical and durable. You can:

• Upload documents, PDFs, decks, and spreadsheets directly
• Import local folders from the desktop app (Mac and Windows)
• Connect to over 400 external tools and apps, including Google Drive, Notion, Linear, Snowflake, and GitHub

Local files stay private to the user who uploaded them, even inside a shared Project.

Brain: memory built around work, not preferences

Most AI memory systems store things about you: your name, your preferences, your writing style. Brain focuses on what the agent did, what worked, what failed, and what corrections were made.

Brain builds a context graph of the work Computer performs. At set intervals, such as overnight, it reviews that graph and updates the agent's working context. Think of it as a nightly debrief: the agent reviews everything it did, updates its internal wiki, and starts the next session already oriented.

Technically, that wiki takes the form of an LLM-readable document automatically loaded into the agent sandbox, giving Computer a map of the user's projects, files, sources, people, and prior sessions. The wiki updates incrementally as Brain reviews sessions, connector results, source document changes, and user corrections.

The feedback loop is recursive. Agents learn which projects, connectors, and sources produce the best outputs. They also learn from mistakes, remembering when a user made a correction or when a source was a dead end. Over time, that means fewer turns, fewer model calls, and better outputs.

The numbers

In Perplexity's internal tests, Brain produced a 25% boost in answer accuracy on repeated tasks, a 16% improvement in recall, and a 13% drop in costs for work that depends on historical context. These are the company's own early metrics, not external benchmarks, and the gains were most pronounced on tasks the agent had already handled. Brain does not make the underlying model smarter. The improvements come from better context retrieval.

Built for teams

Because every task runs from the same context, Projects support multiple concurrent workstreams across a team. When collaborating on a shared Project, every team member works from the same files, context, and skills while personal memory and connectors stay scoped to each person's account.

Inviting someone into a Project does not grant access to their personal memory or connected accounts. Every connector call runs under the individual member's own credentials and permissions. Team members can fork a shared thread to build on work without inheriting the original author's private connectors or memory.

Projects also integrate with communication tools. Type /project in a Computer-connected Slack or Teams channel to bind it to a Project, then kick off and manage work with full Project context from web, mobile, or desktop.

How to set one up

You can create a new Project from scratch or convert an existing session, carrying its files and artifacts with it. From there:

1. Set instructions and defaults — tell Computer how to work in this Project (formats, guardrails like "never publish without asking first") and set the default model for new sessions
2. Turn on Brain — enable Project memory and choose whether it refreshes automatically overnight or on demand
3. Upload or connect files — docs, decks, spreadsheets, and PDFs become durable Project context
4. Connect tools — link Drive, Notion, Linear, Snowflake, and 400+ others
5. Add a Slack or Teams channel — Computer references the channel conversation as Project context

Projects also support Skills, reusable instruction sets for recurring processes. If you find yourself rebuilding the same setup repeatedly, turning it into a Skill eliminates that overhead.

Where it shines, and where it doesn't

Projects with Brain work best for repetitive, context-heavy, multi-session work. The 25% accuracy improvement applies specifically to tasks the agent has seen before, so the more you use it within a domain, the better it gets. Strong fits include:

• Ongoing research and competitive analysis
• Multi-week client deliverables such as reports, decks, and models
• Team workflows pulling from shared data sources like Snowflake or Notion
• Recurring automations with scheduled status updates

For one-off tasks or highly creative work where starting fresh is desirable, the benefits are limited. Developers who need visual feedback loops, live previews, and deep environment control will find it a poor fit. Brain's improvements also don't generalize beyond what the agent has already encountered; broad transfer learning is not the claim here.

A different bet on memory

OpenAI, Google, Mem0, Zep, and Letta are all shipping memory features right now. Perplexity's differentiator is the distinction between user memory and work memory. Most agent frameworks treat memory as personalization. Brain treats it as a performance log. That design philosophy maps more directly to how teams use agents for ongoing work.

Brain is also automatic and inferred rather than explicit. You don't write memories; Perplexity builds the context graph from your sessions and connectors without manual curation. Compare that to Claude's memory system, which gives you more predictability at the cost of more setup. Brain trades control for lower friction.

Signals suggest Brain will power Perplexity Search, the Computer agent, and possibly the Comet browser, positioning memory as a shared connective layer rather than a per-product add-on. Projects is the first surface where that full memory stack comes together in one place.

Availability and pricing

Projects are available to all Computer users now. Existing Spaces migrate automatically. Computer is bundled into the Max plan at $200/month or $2,000/year, which includes unlimited Labs and 10,000 monthly Computer credits. Enterprise Max at $325/seat/month adds heavier agentic and Labs capacity for organizations running Perplexity as core infrastructure. Brain is currently in research preview, so the interface and behavior may still evolve.