Perplexity gives Computer long-term memory

Starting a fresh AI thread and having to re-explain your whole project from scratch is one of modern work's most tedious chores. You correct a tone, hand over background context, and two days later, the chatbot acts like you just met.

Perplexity is trying to eliminate that amnesia. The company just released a sweeping series of updates to its Computer product, headlined by a self-improving memory engine called Brain, a new orchestration model, and the ability to publish web pages right from your prompt.

A memory graph that updates while you sleep

Brain basically turns your past work into an active context system. Instead of dumping every past prompt into an unwieldy context window, Brain builds a private context graph across your chat sessions, connected tools, uploaded files, and previous decisions.

Here is the clever part: it refreshes that graph overnight. The system essentially reviews what succeeded and what failed while you were away, so the next task starts with your preferences already baked in. In testing, Perplexity claims Brain boosted answer correctness by 25% and recall by 16%, while reducing inference costs by 13% on tasks with prior context. Crucially, every stored memory links back to its source session, and you can audit or purge individual items under Customize.

On the model front, Claude Fable 5 is now on board as an orchestrator for heavy, multi-step workflows like deep research. Opus 4.8 also gained a Fast mode for snappier complex drafting on the web. Even better, you can now swap orchestrator models mid-task between turns. You can bring in a heavy-reasoning model to crack a thorny problem, then drop down to a faster, cheaper model to polish the draft without rebooting the thread.

Private market data and instant deployment

Perplexity also beefed up what goes into Computer and what comes out of it.

For inputs, Computer now pulls private company financial intelligence directly from Forge Global. Without needing a separate license or setup, any user can query funding histories, implied valuations, cap table implications, and secondary pricing for private tech giants.

For outputs, Computer can now turn generated sites into live web pages with one click. You can host them on a pplx.app URL or link your Vercel account to push directly to a custom domain. Access controls let you restrict views to yourself, designated teammates, or your whole company, and workspace admins have the power to shut off public publishing altogether to stop accidental leaks.

Rounding out the drop are per-model analytics for tracking credit spend, an account switcher in the Mac desktop app, and a cleaner homepage that tucks verticals like Academic, Health, and Finance into a single menu.

Why it matters

Perplexity is visibly pushing Computer past the boundaries of a clever search bar and into a persistent operating environment for knowledge workers.

For analysts, founders, and developers, this makes Computer well worth testing right now. Having Forge Global private valuations natively accessible saves hours of scouring niche secondary-market databases, while the one-click Vercel publishing turns a research synthesis into a live dashboard in minutes. That said, do not expect production-grade web applications out of the box; these sites are ideal for internal dashboards, rapid prototypes, and client teasers, but serious software still requires traditional engineering pipelines.

Strategically, Perplexity is playing a smart game against rivals who are solely obsessed with shipping bigger base models. The overnight memory refresh is particularly sharp: by processing context asynchronously during downtime, it drives up accuracy and drives down latency during live interactions.

The trade-off is platform lock-in. Once an entire organization relies on a proprietary context graph that gets smarter with every project, leaving Perplexity becomes extraordinarily painful. If you run a privacy-conscious team, the biggest question now is whether the administrative controls are tight enough to keep sensitive institutional knowledge from bleeding across different operational silos.