Image: METAL
Summary
- Perplexity launched Brain on August 26, a self-improving memory system for its agentic Perplexity Computer.
- In online A/B evaluations, turning Brain on raised the overall quality score from 0.53 to 0.62 while cutting token usage by 15%.
- Individual metrics rose as well — correctness up 9.3 points, currentness up 8.0 points, and recall up 8.9 points — a day after the company launched a local version called Portable Computer.
Perplexity gives its agent a "memory"
On August 26, Perplexity unveiled Brain, a self-improving memory system built for its agentic system, Perplexity Computer. Brain works by pulling together conversation sessions, uploaded files, and sources gathered through search, organizing them into a structured knowledge wiki. According to new online evaluation results Perplexity released alongside the launch, turning the feature on pushed the overall quality score up by nearly 9 points while actually cutting token usage by 15%.
Why memory, and why now
If an agent has to reconstruct context from scratch every session, asking about the same topic again means repeating the same search and reasoning work from zero. Brain is designed to compress the conversation history, documents, and sources built up along the way into a single knowledge base that can be pulled up instantly in the next session. Perplexity describes it as a "self-improving" system that continuously folds three types of input — sessions, files, and sources — into that knowledge wiki.
Just a day earlier, on August 25, Perplexity had already announced Portable Computer, a version that runs entirely locally on NVIDIA's DGX Spark with no cloud dependency at all. In that setup, the orchestrator LLM, the sub-agent LLMs, and the entire agent harness all run on the user's own hardware. Brain now takes that a step further, tackling the question of what an agent remembers across sessions — whether it's running locally or in the cloud.
What improved, and by how much
According to the online evaluation chart Perplexity released, comparing Brain on versus off across roughly 854 paired cases over the past 30 days produced the following results:
| Metric | Brain On | Brain Off |
|---|---|---|
| Overall quality score (judged) | 0.62 | 0.53 |
| Relative token usage | 0.85x | 1.00x (baseline) |
The chart carried the annotation "+9pp quality · -15% tokens." Separately, Perplexity broke out the gains by individual metric: correctness improved by 9.3 points, currentness by 8.0 points, and recall by 8.9 points — a three-way breakdown that captures gains the single overall score doesn't fully show on its own.
What's confirmed so far
What's confirmed in this announcement is that Brain is a feature built into Perplexity Computer, that it weaves sessions, files, and sources into a knowledge wiki, and the evaluation numbers reported above. The announcement doesn't specify the exact menu path for toggling Brain on and off, or what the interface looks like for browsing the wiki directly. Portable Computer, the local version announced a day earlier, is reportedly available to Perplexity Pro and Max subscribers and supports PPLX 27B and Qwen 3.8 27B as its runtime models — but this tweet alone doesn't clarify whether Brain is exclusive to that local version or also applies across the cloud version.
Editor's view
Two straight days of Perplexity Computer announcements suggest the company isn't just building a chatbot — it's assembling a full agentic stack, from hardware all the way up to memory. If Portable Computer solved the problem of keeping data from leaving the device by running locally, Brain looks like an attempt to solve the next problem: a local agent that forgets everything the moment a session ends. Set alongside reports that NVIDIA is in talks to invest in Perplexity at a valuation above $30 billion, the picture sharpens — this is a company aiming to own the entire stack of an agentic computer market for individuals and businesses, well beyond search.
Anyone who's worked with agentic products knows token cost ultimately drives the underlying cost of the service. If the numbers hold up — a 9-point quality gain alongside a 15% cut in tokens — that's a change that hits the operating cost structure harder than it hits the user-facing experience. Teams evaluating agentic tools should weigh how memory and context-compression approaches like this actually move the bill, rather than just comparing model performance charts.
In the coming weeks, expect competing agentic products to roll out similarly named "memory" features of their own. Compressing cross-session memory into a structured knowledge base is quickly becoming a standard component that any serious agentic product will need to have.
Comments