Perplexity launched Brain on June 18, a memory system for its Computer agent that tracks what the agent did, which sources held up, what corrections the user made, and what failed, then uses that history to start the next session with full context instead of from scratch.
“With Brain, Computer starts each task with full context of your projects, decisions, and sources instead of from scratch,” Perplexity said, according to Decrypt. “Every memory links back to the session, file, or source it came from with full transparency and control.”
How It Works
Every time Computer completes a task, Brain adds it to a context graph. The graph tracks which connectors were used, which sources returned reliable information, what corrections the user applied, and what produced bad results. At set intervals (overnight by default), Brain synthesizes the graph and updates a personal LLM wiki that loads into Computer’s sandbox before the next task begins. Each memory entry links back to its originating session or file.
Perplexity’s internal metrics show Brain improves answer correctness by 25% on tasks Computer has previously handled, improves recall by 16%, and cuts the cost of context-heavy tasks by 13%, per Decrypt. These are internal numbers, not third-party benchmarks.
The logic is that most AI memory systems focus on the user: preferences, habits, personal details. Brain’s memory focuses on the work: what the agent tried, what succeeded, what source led somewhere useful. An agent that starts each morning knowing which sources failed last week wastes fewer tokens rediscovering that information.
Existing Approaches
Perplexity is not the first to ship agent memory. OpenClaw has been doing versions of this for months using markdown files and a SQLite database with FTS5 full-text search to persist context across sessions, Decrypt noted. OpenClaw added “providence labels” in April 2026, tagging each stored memory as observed, user-confirmed, model-inferred, or imported from a transcript, so the agent knows how reliable any given fact is.
Nous Research’s Hermes agent goes further. After each completed task, Hermes evaluates the outcome, extracts reusable reasoning patterns, and writes them as skill files in plain markdown. The next time it encounters a similar problem, it loads the skill instead of reasoning from scratch.
Both OpenClaw and Hermes are self-hosted. Users run them on their own hardware, and data stays local. Brain runs entirely on Perplexity’s infrastructure. Users get transparency into what is stored but not ownership of the data. The distinction matters for teams with data sovereignty requirements.
Pricing and Availability
Brain is rolling out in Research Preview for Max ($200/month) and Enterprise Max subscribers. Memories are accessible under “Customize” in the sidebar. Perplexity said new capabilities are coming soon, with no timeline attached.
The Cost Equation
The 13% token cost reduction on context-heavy tasks could be significant at scale. For teams running recurring research workflows, competitive monitoring, or weekly reporting, the compounding effect of an agent that stops reinventing context every session is measurable in both time and compute spend. Whether that reduction justifies the $200/month subscription depends on task volume and how much of the work is genuinely repetitive.