Aravind Srinivas (@AravSrinivas), Perplexity's co-founder and CEO, announced Projects on Perplexity Computer on July 30, 2026: "turning Computer into a multiplayer agentic operating system for work with persistent memory, files, and sessions scoped across hubs and users. Available to all users!"
It's the collaboration layer Perplexity's existing agent stack was missing — Computer Brain already gave individual users a self-improving, persistent context graph; Projects extends that same self-improvement loop to teams working inside a shared, bounded context rather than isolated single-user sessions.
TL;DR
| Question | Answer |
|---|---|
| What launched? | Projects on Perplexity Computer |
| Announced | July 30, 2026, @AravSrinivas |
| Framing | "Multiplayer agentic operating system for work" |
| Scoped by | Hubs and users — persistent memory, files, and sessions |
| File system | Shared — single source of truth for all collaborators on a Project |
| Integrations | Google Workspace, Slack |
| Customization | Support for custom skills and instructions per Project |
| Memory engine | Computer Brain — self-improvement loops now run per Project, across all collaborators |
| Availability | All users, per the announcement |
What "multiplayer agentic operating system" actually means here
Srinivas's own phrase is doing a lot of work, so it's worth unpacking against what's actually described in the thread rather than taking the framing at face value. Three concrete components were named across the announcement:
- Persistent memory, files, and sessions — scoped across hubs and users. Not per-user isolation; a Project is a shared container multiple people and multiple Computer sessions operate inside simultaneously.
- A shared file system. Described explicitly as "a single source of truth for all collaborators" — the same files, visible and referenceable by everyone in the Project, rather than each person's Computer session maintaining its own separate file context.
- Integrations to Google Workspace and Slack, plus custom skills and instructions. This is the connective tissue to where teams already work — Projects isn't asking teams to migrate their existing docs and chat into a new silo, it's meant to sit alongside Workspace and Slack, pulling context from and pushing work into tools already in use.
Put together, "operating system" is the right register for what's being described: not a single-purpose feature, but a substrate multiple people, files, integrations, and Computer sessions all run on top of, with a shared state layer underneath all of it.
Brain's self-improvement loop, extended to teams
The most technically interesting line in the announcement is easy to miss: "every project goes through self-improvement loops on its memory, leading to better contextual understanding over time, thanks to offline compute that Brain performs between tasks and sessions of the collaborators on a Project."
That's a direct extension of Computer Brain, which explainx.ai covered at its original June 19, 2026 launch — a continuously learning context graph that updates overnight and feeds into every subsequent task, delivering (per Perplexity's own benchmarks at the time) a 25% accuracy improvement, 16% better recall, and 13% cheaper per-task cost on tasks requiring past context.
The original Brain was per-user: your sessions, your connectors, your files, consolidated overnight into your context graph. Projects generalizes that same mechanism to a shared, multi-collaborator scope — Brain's offline compute now runs against the combined activity of everyone working inside a given Project, not just one person's history. In principle, that means a Project's context graph should compound faster and cover more ground than any single collaborator's individual Brain memory would on its own, since it's drawing on the full team's sessions rather than one person's.
Whether that scales cleanly — more collaborators contributing more raw session data doesn't automatically mean cleaner or more useful consolidated memory — is an open question this launch doesn't answer on its own. Team-scoped memory systems generally face a real tradeoff between breadth of context and noise from irrelevant individual tangents; how well Brain's consolidation handles that tradeoff at Project scale is worth watching for, not assumed from the announcement alone.
Where Projects sits relative to Perplexity's other recent Computer work
Perplexity has been shipping fast on the Computer agent stack across July 2026. Projects is the newest piece, but it's not the only recent one:
| Update | Date | What it addresses |
|---|---|---|
| Computer Brain | Jun 19, 2026 | Per-user persistent memory — the original self-improving context graph |
| GLM 5.2 orchestrator | Jul 9, 2026 | Which model routes/executes Computer tasks — 0.344x Opus cost, native advisor escalation |
| Projects | Jul 30, 2026 | Collaboration and memory-scoping layer — multi-user, multi-session, shared files |
These are complementary, not competing systems. The GLM 5.2 orchestrator is about what runs a task and how much it costs; Projects is about the shared context that task runs inside and who else has access to it. A given Project could, in principle, use any Computer backend — including the GLM 5.2 orchestrator — for actual task execution, while Projects handles the surrounding scope: files, memory, integrations, and who's in the room.
What's genuinely new vs. what's a rebrand of existing pieces
Worth being precise about which parts of this announcement are net-new capability versus existing Perplexity infrastructure repackaged under a new name:
| Component | Status |
|---|---|
| Persistent context graph / self-improving memory | Existing — this is Brain, now applied at Project scope |
| Multi-user collaboration on a shared agent session | New — the "multiplayer" framing is the headline addition |
| Shared file system across collaborators | New, as a described feature — no prior explainx.ai coverage of file-sharing inside Computer |
| Google Workspace / Slack integrations | Partially existing — Perplexity has had connector integrations before; scoping them per-Project is the new part |
| Custom skills and instructions per Project | New, as a project-scoped concept — consistent with the broader 2026 trend toward agent skills as a packaging unit for reusable agent behavior |
The honest read: the memory engine (Brain) isn't new — the collaboration and scoping model wrapped around it is. That's a meaningful ship, but it's an evolution of Perplexity's existing agent architecture rather than an entirely new system built from scratch.
Practical questions worth checking before adopting Projects for a team
- Access tier — Brain's original research preview was Max-subscriber-only ($200/month); confirm whether Projects requires the same tier despite the "available to all users" framing, since announcement copy and actual account-level gating don't always match on day one of a launch.
- Data boundaries — a shared file system and cross-collaborator memory consolidation raises real questions about what's visible to whom inside a Project, and whether that scope is cleanly separated from a user's personal (non-Project) Computer memory. Perplexity's Brain documentation previously emphasized full transparency and source traceability for individual memory; verify the same standard extends to multi-user Project memory.
- Integration depth — "integrations to Google Workspace and Slack" covers a wide range of possible depth, from read-only context pulling to two-way write access. Test with a low-stakes Project before connecting a team's primary Workspace or Slack instance.
- Custom skills scope — confirm whether Project-level custom skills are private to that Project or shareable/reusable across a user's other Projects, since that materially affects how much setup work is duplicated per new Project.
- Offboarding and access changes — a shared file system and pooled team memory raise a question standalone per-user Brain never had to answer: what happens to a Project's shared context when a collaborator leaves the team or loses access mid-project? Confirm whether removing a user strips their contributed sessions from the consolidated memory or simply revokes their future access while their historical contributions remain baked into the graph.
How this compares to team features in Claude and ChatGPT
Perplexity is not the first AI product to ship team-scoped context. Anthropic's Claude for Work and OpenAI's ChatGPT Enterprise both offer shared workspaces, and Claude Cowork specifically targets collaborative agent sessions. The differentiator Projects is claiming is not "teams can use this together" — that bar has been cleared industry-wide — but persistent, self-improving shared memory that compounds across a team's activity automatically, without anyone manually curating what gets remembered. Most competing team features handle shared documents and shared chat history well; fewer claim an offline consolidation process that actively gets better at a specific project's context over time, the way Brain's nightly loop does. Whether that compounding advantage holds up under real team usage, with its noisier and more varied inputs than a single user's session history, is the open question worth revisiting once early adopters report back.
Summary
Perplexity launched Projects on Computer on July 30, 2026 — a collaboration layer turning the agent into what CEO Aravind Srinivas calls a "multiplayer agentic operating system for work." Projects scope persistent memory, files, and sessions across hubs and users, add a shared file system as a single source of truth for collaborators, integrate with Google Workspace and Slack, and support custom skills and instructions per Project. Under the hood, Computer Brain's self-improvement loops — previously per-user — now run against the combined activity of every collaborator on a Project. It's presented as available to all users, extending rather than replacing the memory architecture Perplexity shipped in June and the GLM 5.2 orchestrator work from earlier in July.
Related on explainx.ai
- Perplexity Brain — the self-improving memory system inside Computer
- Perplexity Computer GLM 5.2 orchestrator — 0.34x Opus cost
- Perplexity Search-as-Code — agentic search
- Perplexity Wandr benchmark — open-source research agents
- Agent skills — complete guide
- What is MCP? Model Context Protocol guide
- What is an Obsidian vault as AI agent memory?
Source: @AravSrinivas — July 30, 2026
Feature details reflect the public announcement as of July 31, 2026. Access tiers, integration depth, and pricing can change quickly after launch — verify current availability directly in your Perplexity account before rolling Projects out to a team.