이미지: @perplexity_ai (X) 화면 갈무리 · METAL LAB 편집
Summary
- Perplexity launched Brain, a self-improving memory system for its Perplexity Computer agent, on August 26.
- 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%.
- Breakdown metrics showed accuracy up 9.3 points, currentness up 8.0 points, and recall up 8.9 points; a day earlier, Perplexity had released a local version called Portable Computer.
- 발표
- Perplexity, 2026-08-26 (X)
- 제품명
- Brain — Perplexity Computer용 자가개선 메모리 시스템
- 핵심 기능
- 세션·파일·출처를 구조화된 지식 위키로 컴파일
- 평가 방식
- 페어드 A/B, 최근 30일, 페어당 약 854개 사례
- 종합 품질 점수
- Brain 켬 0.62 vs 끔 0.53 (+9pp)
- 상대 토큰 사용량
- Brain 켬 0.85배 vs 끔 1.00배 (-15%)
- 세부 개선폭
- 정확성 +9.3점 · 최신성 +8.0점 · 회상력 +8.9점
- 관련 제품
- Perplexity Computer (로컬 버전 Portable Computer는 8월 25일 DGX Spark서 공개)
Perplexity gives its agent a "memory"
On August 26, Perplexity introduced Brain, a self-improving memory system that plugs into its agent system, Perplexity Computer. Brain pulls together conversation sessions, uploaded files, and sources gathered through search, then organizes them into a structured knowledge wiki. According to new online evaluation results Perplexity released alongside the launch, turning the feature on lifted the overall quality score by nearly 9 percentage points while token usage actually dropped 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 immediately in the next session. Perplexity describes it as a "self-improving" system that continuously folds three kinds of input — sessions, files, and sources — into that knowledge wiki.
Just a day earlier, on August 25, Perplexity had also unveiled Portable Computer, a version that runs entirely locally on NVIDIA's DGX Spark with no cloud dependency. In that setup, the orchestrator LLM, sub-agent LLMs, and the entire agent harness all run on the user's own hardware. Brain represents the next step in that direction — 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 published, comparing Brain-on and Brain-off states 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 improvements across three specific metrics: correctness rose 9.3 points, currentness rose 8.0 points, and recall rose 8.9 points. That breakdown captures three distinct gains that a single composite score wouldn't reveal on its own.
What's confirmed so far
What's confirmed by this announcement: Brain is a feature built into Perplexity Computer, it stitches sessions, files, and sources into a knowledge wiki, and these are the evaluation numbers behind it. What the announcement doesn't cover is the specific menu path for turning Brain on or off, or what a screen for browsing the wiki directly might look like. Portable Computer, the local version released the day before, is reportedly available to Perplexity Pro and Max subscribers and supports PPLX 27B and Qwen 3.8 27B as its runtime models. Whether Brain applies only to that local version or extends across Perplexity's cloud offerings as well isn't clear from this announcement alone.
Editor's take
Two consecutive days of Perplexity Computer announcements suggest the company isn't just building a chatbot — it's assembling a full agent stack, from hardware 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 negotiating an investment in Perplexity at a valuation north of $30 billion, the picture comes into sharper focus — this is a company aiming to own the entire agent-computer stack for personal and enterprise use, not just search.
Anyone who's worked with agentic products knows that token cost ultimately drives service economics. If the numbers hold up — a 9-point quality gain alongside a 15% cut in tokens — that's a change that matters more for operating cost structure than for how users experience the product day to day. Teams in Korea looking to adopt agentic tools would do well to look past model performance leaderboards and ask how memory and context-compression approaches like this one actually show up on the bill.
Expect competing agent products to roll out similarly named "memory" features in the coming weeks. Compressing cross-session memory into a structured knowledge base is fast becoming a standard component that any serious agent product will need to have.
Comments