On June 18, 2026, Perplexity introduced "Brain" — a self-improving memory system for its "Computer" agent product. Unlike traditional memory, which only stores user preferences, Brain builds a living "context map" of what the agent has actually done — then periodically, overnight, synthesizes it into a kind of "wiki" so the agent can teach itself. The feature is in Research Preview, available only to the Perplexity Max and Enterprise Max plans.
Quick summary
- When: June 18, 2026.
- What: "Brain" — self-improving memory for Perplexity's "Computer" agent.
- Scope: Research Preview, only for the Perplexity Max & Enterprise Max plans.
- How it works: Builds a living "context graph" → overnight synthesizes it into an "LLM wiki" → loads it into the sandbox so the agent can improve itself.
- Numbers (Perplexity internal): +25% accuracy, +16% recall, −13% cost per task requiring historical context.
What is Brain?
According to Perplexity's announcement, Brain is a new memory layer attached to the "Computer" agent product. The core difference: ordinary AI memory mostly records the preferences of the user (for example, "prefers concise answers" or "uses the metric system"). Brain goes further — it builds a living "context graph" that records the very work the agent has done: what worked, what failed, and how errors were fixed.
The feature is in the Research Preview stage — an early test build — and is only available to users on the Perplexity Max and Enterprise Max plans.
How Brain works
Perplexity describes the mechanism as having two parts. First, while the agent is working, Brain continuously updates the context graph — a network of context reflecting the agent's actions, outcomes, and error handling. Second, periodically overnight, the system synthesizes what it has accumulated into a kind of "LLM wiki" — a condensed knowledge base — and loads it back into the agent's sandbox. As a result, the agent essentially "teaches itself" from experience instead of starting from scratch every session.
The numbers Perplexity published
Perplexity says Brain delivers improvements on tasks that require historical context. According to Perplexity's internal numbers (no independent benchmark):
- +25% accuracy.
- +16% recall (the ability to recall the right relevant information).
- −13% cost for each task that requires historical context.
It is worth emphasizing that these are according to Perplexity (internal numbers, no independent benchmark) — so they should be treated as the company's own claims, not yet verified by a third party.
Source traceability
One notable point on transparency: according to Perplexity, every item in Brain's memory is linked back to its origin. In other words, when the agent uses a piece of learned "knowledge," the system can trace that piece back to the action or event it came from — making it easier to audit and explain than an opaque block of memory.
A perspective for Vietnamese businesses
Brain's direction reveals a clear trend: AI agents increasingly accumulate context over time — not just preferences, but how your organization operates, which processes work, and which errors recur. And that is precisely the issue: an organization's memory and context are sensitive assets.
When this memory layer sits on a foreign provider's infrastructure, a business faces a double risk: a compliance risk (PDPL, cross-border data transfers) and the risk that the organization's "operational memory" is beyond its control. Namtech's view: for core data and processes, the memory/context layer should be kept inside the company's own internal systems — so you can harness the power of self-improving agents while not handing your knowledge assets to an outside party.
FAQ
Can I use Brain right now?
At the time of writing (June 22, 2026), Brain is in Research Preview and is only open to the Perplexity Max and Enterprise Max plans. This is reference information per Perplexity's announcement and may change.
Are the +25% / +16% / −13% numbers trustworthy?
These are Perplexity's internal numbers, published by the company itself, with no independent benchmark to verify them. They should be treated as vendor claims, not results confirmed by a third party.
How is Brain different from ordinary AI memory?
Ordinary memory mostly stores user preferences. Brain builds a living "context graph" of what the agent has done (success/failure/error-fixes), then overnight synthesizes it into an "LLM wiki" loaded into the sandbox so the agent can improve itself.
Keep your organization's "memory" in your own hands
An organization's memory and context are sensitive assets. Namtech deploys internal AI platforms — agents and the memory layer run on your own infrastructure, so your data and knowledge never leave the organization.
Book a free consultationNote: This article is compiled from public sources as of June 22, 2026; the performance numbers are Perplexity's internal figures, with no independent benchmark. For reference only — not technical or legal advice.