Model: Perplexity · Brain
Brain is a self-improving memory system that lets Perplexity’s Computer agent build a private context graph from your sessions, files and past decisions, then refresh it overnight so each task starts knowing what worked. It shipped in Perplexity’s 13 July 2026 release, and it matters because AI answers no longer start from a blank slate.
Source: Perplexity changelog — Brain, faster Computer models, and website publishing (13 July 2026)
- On 13 July 2026 Perplexity shipped Brain — Computer now learns from every task, building a private context graph across sessions, connectors and files, then refreshing it overnight.
- Perplexity reports Brain lifts answer correctness by 25% and recall by 16%, while cutting cost on context-heavy work by 13%.
- The same release added faster Computer models (Opus 4.8 in Fast mode), mid-task model switching, and website publishing to a .pplx.app URL or your own domain via Vercel.
- Computer can now research private companies — fundraising history, valuations and secondary pricing — using Forge Global data, for all Computer users.
- Memory is the headline for brands: an agent that remembers what worked keeps reaching for the same sources, so presence inside an account compounds.
What Perplexity shipped on 13 July
Perplexity’s 13 July 2026 changelog entry, “Brain, faster Computer models, and website publishing”, bundles five changes to Computer, its agent product. In Perplexity’s own one-line summary, Computer “now remembers past work, responds faster, switches models mid-task, publishes websites, and researches private companies.”
Bundled release notes are easy to read as housekeeping, and most of a changelog usually is. This one is not. Three of those five items change where brands can surface and how consistently they get picked — a bigger shift than the Deep Research upgrade we logged in June.
What is Brain and how does it work?
Brain is memory for the agent, not memory of you. Perplexity describes a system where Computer “learns from every task”, building “a private context graph across your sessions, connectors, files, and past decisions”, then refreshing it overnight “so each new task starts already knowing what worked, what failed, and how you like things done.”
The distinction is the interesting part. Most memory features ask you to save preferences, then recall them — the pattern we covered in ChatGPT memory sources. Brain is automatic and inferred: you never write the memories, Perplexity builds the graph from what the agent actually did. Perplexity reports it improves answer correctness on repetitive tasks by 25%, recall by 16%, and cuts cost on context-heavy work by 13%. It reached Max and Enterprise Max subscribers first in research preview, with wider access following.
Why does agent memory compound?
An agent with memory is not a fresh audition every time. If Computer researched your category last week and one source proved useful, that result is in the graph — and the overnight refresh reinforces it. Being useful once makes you more likely to be reached for again. Being absent compounds the same way.
One honest caveat: the graph is private and per-account. Brain does not change what a stranger sees on their first ever query, and you should not read it as a global ranking signal. What it does change is the repeat case — a procurement team, an analyst or an agency running your category through Computer for weeks. There, the first impression stops being disposable. That is a different mechanism from the source-trust logic in how LLMs decide which brands to cite, and it sits on top of it, not instead of it. The groundwork in building a GEO strategy and is your brand ready for AI search? still decides whether you make the graph at all — as does earning the kind of source status we traced in Google preferred sources.
It is worth noting the opposite pattern exists too: ChatGPT built Fast Answers to bypass memory for speed. The platforms are actively split on whether memory helps or hinders an answer, which is exactly why it is worth measuring rather than assuming.
What changed in the Computer models?
Two things. Opus 4.8 is now available in Fast mode for quicker responses on complex work, and Claude Fable 5 joins as an orchestration option for involved research that needs stronger planning. Both are models we have covered on launch — Opus 4.8 and Fable 5 — now doing orchestration duty inside someone else’s agent.
The subtler change is mid-task model switching: you can change the orchestrator model between follow-up turns inside a single Computer task, matching cost to complexity without restarting. Which model orchestrates affects which sources get pulled and which brands get named — the effect we measured when ChatGPT swapped its fallback to GPT-5.5 Instant Mini, when GPT-5.2 was retired, and when the model picker was simplified. Now that swap can happen within one task, which makes a single answer a weaker sample than it already was. Claude Sonnet 5 made the same point from the model side.
Can Perplexity publish websites now?
Yes. Computer can build a website or full-stack app, preview it, and publish it live in a click — to a persistent .pplx.app address with no hosting, DNS, backend or environment configuration, or to your own custom domain by connecting Vercel. Access is scoped: just you, specific people, your whole organisation, or everyone on the web. Organisations can switch public publishing off entirely and keep it internal.
That is the second AI publishing surface to open in under a week. OpenAI opened ChatGPT Sites in public beta on 9 July, alongside ChatGPT Work. The answer engines are becoming publishers, and each new AI-built page competes for the same attention as the human-made ones — the wider pattern in this week in AI and in Claude arriving in Microsoft 365 Copilot.
Why private company research matters
This is the item most likely to be overlooked, and it may be the one that touches the most brands. Computer can now run financial research on private companies — fundraising history, valuations and secondary pricing — sourced from Forge Global, available to every Computer user with no extra setup.
Most B2B brands are private companies. Until now, an AI answer about your funding or valuation was thin or refused. Now there is a sourced answer whether or not you participated in producing it, and it will be repeated to investors, candidates and buyers. If the underlying record is stale, an AI answer will confidently say so — the risk we set out in the accuracy imperative, and the same quiet reshaping we described in how geopolitical events reshape AI visibility.
Five changes in one Perplexity release
Perplexity changelog, 13 July 2026. Highlighted items change where brands surface or how consistently they are picked.
How to check your Perplexity presence
Memory makes one-off checks less meaningful, because the thing you are measuring now changes with use. A practical way to test what this release does to your category:
- Ask about your own company. On any Computer plan, request its funding history and valuation. Check what Forge Global data says about you, and whether it is current.
- Run your category cold. Use a fresh account with no history and note which brands Computer names. That is your blank-slate baseline.
- Then run it warm. Repeat the same brief in an account that has researched the category for a fortnight, and compare. The gap is the memory effect.
- Switch the orchestrator. Re-run one brief on Opus 4.8 Fast and again on Fable 5. If the named brands change, a single answer was never your number.
- Watch pplx.app pages. Search your category for published Computer sites — they signal where AI-built pages are entering your space.
- Re-check monthly. Brain refreshes overnight and access is still widening, so treat any single reading as a snapshot, not a verdict.
How reconnAI tracks Perplexity
reconnAI monitors how the leading AI models — ChatGPT, Claude, Gemini, Perplexity, Copilot and Google AI Overview — answer questions across regions, and re-baselines whenever a platform changes how its answers are produced. A release that gives an agent persistent memory, a new publishing surface and a private company data source is exactly that kind of change, and it is why we measure trends rather than single answers.
If you want to see how Perplexity represents your brand now that its agent remembers, get in touch with our team or explore how AI visibility tracking works. If your interest is the commercial side of AI answers, our overview of ChatGPT advertising covers where paid placement is heading.
Perplexity Brain: your questions answered
What is Perplexity Brain? Brain is a self-improving memory system for Perplexity’s Computer agent. It builds a private context graph across your sessions, connectors, files and past decisions, then refreshes it overnight so each new task starts already knowing what worked, what failed and how you like things done.
When did Brain launch? It appeared in Perplexity’s 13 July 2026 changelog entry alongside faster Computer models and website publishing, having reached Max and Enterprise Max subscribers in research preview first.
How is Brain different from ChatGPT memory? ChatGPT memory largely remembers you and your stated preferences. Brain remembers what the agent did — it is automatic and inferred rather than manually curated, and it self-improves on a scheduled refresh.
Does Brain change which brands Perplexity names? Within an account, plausibly yes over time — sources that proved useful are reinforced in that account’s graph. It is private per account, so it does not alter what a first-time user sees. Testing a cold account against a warm one is how you tell.
Can Perplexity publish a website? Yes. Computer can publish a site or app to a persistent .pplx.app URL with no hosting setup, or to a custom domain via Vercel. You choose who can see it, and organisations can turn public publishing off.
About reconnAI
reconnAI tracks how the major AI models represent topics and sources across ChatGPT, Claude, Gemini, Perplexity, Copilot and Google AI Overview — across multiple regions. We monitor how those models answer and how they change over time, so you can stay ahead of shifts in the AI landscape.
Last updated: 17 July 2026.