explainx / blog
Perplexity launched Brain—a continuously learning context graph for its Computer agent. 25% accuracy improvement, 16% better recall, 13% cheaper per task. Here's what it is and how it works.
Jul 9, 2026
Perplexity post-trained GLM 5.2 for the Computer harness — research preview with an advisor tool that escalates to stronger models, ~half Opus cost on WANDR, hosted on US Nvidia B200s. explainx.ai breaks down the July 9 orchestrator drop.
Jun 29, 2026
video-use is an open-source skill for Claude Code (and Codex, Hermes, Openclaw) that edits videos via natural language — no timeline scrubbing, no NLE menus. It reads footage as transcript text, reasons over word-level timestamps, calls ffmpeg, self-evaluates every cut, and outputs final.mp4. 11.6k GitHub stars in two months. Here is the full setup and how it works.
Jun 28, 2026
Four serious tools now compete for the same territory: no-code and low-code AI workflow building. Langflow, n8n, Make, and Flowise are not interchangeable. This comparison breaks down what each tool actually does well, where each one falls short, and gives you a three-question framework to pick the right one for what you are building in 2026.
Perplexity AI shipped a feature last night that most AI labs have been dancing around for two years: a memory system that actually compounds. Brain is a continuously learning context graph built into Perplexity's Computer agent—and unlike most "memory" features, it updates itself overnight and feeds automatically into every subsequent task.
Perplexity CEO Aravind Srinivas announced it directly: "Brain is a self-improving context-graph of all your sessions, connectors, and files. Brain updates itself overnight with fresh context proactively, and feeds itself to every task on Computer, allowing Computer to be stateful and self-improving."
Update — July 9, 2026: Perplexity previewed a GLM 5.2 orchestrator for Computer — post-trained for the harness, 0.344× Opus cost, native advisor escalation, US B200 hosting.
Update — July 15, 2026: Claude Memory Heist — Ayush Paul showed claude.ai's default memory plus web_fetch browsing could silently exfiltrate PII; contrast with Brain's connector-scoped context graph and source traceability.
Most AI agents start each session cold. They have no knowledge of what you worked on yesterday, what decisions were made last week, or what context matters for your current project. Brain is Perplexity's answer to that problem.
The architecture:
Context graph — Rather than a flat list of "memories," Brain builds a structured graph of your projects, decisions, files, and sources. Relationships between items are preserved—a decision made in one session can connect to a file from a connector and a result from a prior search.
Overnight updates — The context graph refreshes nightly, pulling in new sessions, connector data, and file changes. You don't manage it manually; it grows with your use.
Automatic injection — When you start a task in Computer, Brain feeds the relevant context from the graph into the task automatically. You don't have to paste prior context or re-explain your situation.
Source traceability — Every memory in Brain links back to the specific session, file, or connector it came from. You can inspect where any given piece of context originated and remove it if you want.
Perplexity released internal benchmark results comparing Computer tasks with Brain enabled vs. without it:
| Metric | Improvement with Brain |
|---|---|
| Answer correctness (context-dependent tasks) | +25% |
| Recall | +16% |
| Cost per task | -13% |
The cost reduction is notable—Brain doesn't just improve accuracy, it makes Computer more efficient. Because the agent starts with context instead of having to re-derive it, it needs fewer steps and fewer tokens to reach the right answer.
The gains are specifically on tasks that require past context. For tasks that are entirely new and self-contained, Brain has less impact. The compound effect shows up over time, as your context graph grows richer with each session.
Brain is available as a research preview for all Perplexity Max subscribers ($200/month).
To access your saved memories:
You can view, inspect, and delete individual memories from this interface. Each memory shows its source—which session, file, or connector it came from.
The Perplexity team confirmed that Brain is a research preview, meaning the interface and behavior may change as they iterate based on feedback.
AI memory has been a hot topic in 2026, with several different approaches emerging. Brain sits in a specific part of the design space.
Claude's memory works through two mechanisms: claude.md files that carry persistent instructions across sessions, and Projects—which let users store context that persists within a named project. Both are explicit and user-managed: you write the memory, you manage it, you update it.
Brain is different: it's automatic and inferred. You don't write memories; Perplexity builds the context graph from your sessions and connectors without manual curation.
Trade-off: Explicit memory (Claude) gives more control and predictability. Automatic memory (Brain) is lower friction but requires trusting the system to infer correctly.
ChatGPT's memory feature stores facts and preferences from conversations, displayed as a list of remembered items that users can edit or delete. It's user-facing and relatively flat—there's no structured graph, just a list.
Brain's context graph is more structured: relationships between items are preserved, and the overnight batch update means it integrates across connectors and files, not just chat history.
One X user (Dan, @Daniel_Vogler) noted: "Already built this myself locally with Opus 4.8." That's the realistic benchmark—Brain is doing something that power users can already replicate with custom tooling. The value Perplexity is adding is packaging it into a zero-configuration system that works for non-engineers.
The reason Brain matters goes beyond Perplexity's product specifically. It's about a structural limitation in how most AI agents work today.
Most agents are stateless by default. Each session begins with no knowledge of prior sessions. The user provides context at the start of every task, which is repetitive, error-prone, and limits how sophisticated the agent's behavior can become over time.
The dream of "self-improving" agents requires overcoming this. An agent that knows what approaches worked, what sources you trust, what decisions were made in related projects, and what terminology you use in your domain is qualitatively more useful than one that starts from zero every time.
Brain's overnight update model is one specific approach to this problem:
The limitation of this model is latency: information from today's session isn't available to tonight's task if the batch runs at a fixed overnight schedule. Real-time streaming memory—where the context updates mid-session as you work—is a harder problem that none of the major players have fully shipped yet.
Brain is available to Max subscribers. At $200/month, Perplexity Max is positioned against enterprise-tier AI tools, not consumer subscriptions.
The performance improvements—25% accuracy, 16% recall, 13% cost savings—are framed in those terms. For teams running Computer on repetitive research, coding, or data analysis tasks, a 13% cost reduction compounds meaningfully at scale. A 25% accuracy improvement on context-dependent tasks can mean the difference between usable output and output that requires significant correction.
Whether Brain justifies the Max subscription price depends on how heavily you use Computer and how much your tasks depend on prior context. For occasional users, the overnight update cadence may not accumulate enough context to show meaningful benefits. For daily power users with connected data sources and ongoing projects, the compound effect is the point.
The response on X has been positive but measured. Key themes:
Enthusiasm for the concept: Users who've been frustrated by stateless AI agents see Brain as a step toward the self-improving loop they've wanted.
Skepticism about accessibility: Multiple users pointed out the $200/month price as a barrier. One user (@QtheArsenal) framed the competitive tension directly: "In the world of Codex and Claude, why would I sign up blind?"—capturing the challenge Perplexity faces competing against deeply entrenched alternatives.
DIY comparison: Power users noting they've built equivalent systems locally underscore that the technical approach isn't novel—the packaging is.
The overnight update question: Some developers flagged the batch-update cadence as a limitation for real-time workflows. The gap between "I just made this decision" and "Brain knows about it" is meaningful for fast-moving projects.
Brain is a research preview, which means Perplexity is actively iterating on it. The specific behaviors worth watching:
Update frequency: Will overnight batches move toward real-time? Or will Perplexity introduce manual memory triggers for time-sensitive context?
Connector expansion: Brain currently integrates sessions, connectors, and files. As Perplexity adds more connector types (calendar, project management tools, code repos), the context graph gets richer.
Privacy controls: Source traceability is a good start, but enterprise users will want more granular controls over what Brain is allowed to retain—particularly around sensitive project data.
Pricing for lower tiers: If Brain proves its value at Max, expect Perplexity to evaluate bringing some version of it to lower tiers as a retention lever.
What Perplexity is building with Brain is a particular vision of what an AI agent should be: not a powerful-but-amnesiac tool you reset every session, but a progressively more capable colleague that gets better at helping you specifically as it learns more about your work.
The 25% accuracy improvement on context-dependent tasks is the clearest signal that the approach is working. The 13% cost reduction suggests the efficiency gains are real, not just cosmetic.
Whether that's enough to differentiate in a market where Claude, GPT, and Gemini are all shipping memory features of their own is the question Perplexity is betting on. The answer will depend on whether users are willing to pay $200/month for an agent that specifically compounds over time—or whether they prefer the flexibility of building their own memory systems on top of cheaper foundation models.
Brain is a real product answering a real problem. The research preview label is honest: there's more to build. But the direction is clear, and the early results are worth paying attention to.
Perplexity Brain launched on June 18, 2026 as a research preview for Max subscribers. Access it under Customize in the Perplexity sidebar. For the broader context on AI agent memory architectures, see our coverage of Claude Fable 5's long-context and memory capabilities. July 15, 2026: Ayush Paul's Memory Heist shows why persistent assistant memory plus web browsing is a high-value attack surface — relevant for any product shipping Brain-style retention.
