Most AI memory features are built around the user: your name, your preferences, your writing style. Brain takes a different angle. It remembers the work — what the agent tried, what got corrected, and which sources turned out to be reliable. Every time Computer finishes a task, Brain logs it into a context graph. At set intervals, by default overnight, that graph gets synthesized into a wiki-style memory store that loads into the agent's sandbox before its next run.
"With Brain, Computer starts each task with full context of your projects, decisions, and sources instead of from scratch," Perplexity said in its announcement. Each memory entry links back to the original session, file, or source, giving users a way to trace any decision the agent made back to its origin — a transparency feature aimed at addressing the "black box" criticism often leveled at autonomous agents.
- +25% answer correctness on tasks Computer has handled before
- +16% improvement in recall accuracy
- -13% reduction in cost for context-heavy tasks
- $200/month — Perplexity Max subscription tier required for access
These are Perplexity's own internal figures, not independently verified benchmarks, but the underlying logic holds up: an agent that already knows which sources failed last week doesn't need to burn tokens rediscovering that failure. For recurring workflows — competitive monitoring, weekly reports, multi-step research — that compounding efficiency could matter more than raw model capability.
The key difference is where the data lives. OpenClaw and Hermes are self-hosted — memory stays on hardware the user controls. Brain's context graph, by contrast, lives entirely inside Perplexity's infrastructure. Users get visibility into what's stored, but not ownership of it. That distinction matters for any team weighing data sovereignty against convenience.
It's also worth being precise about what "self-improving" means here. Brain makes Computer faster and more accurate at tasks it has already done for a specific user — it does not make the underlying model smarter, and it does not generalize across unrelated domains. Knowledge gained helping with a financial research task won't transfer to a coding task. That cross-domain generalization problem remains open, and Brain doesn't claim to solve it.
- Brain logs agent actions — not user preferences — into a context graph that's synthesized into memory overnight.
- Early internal metrics show meaningful gains in correctness, recall, and cost for repeated tasks.
- Memory lives on Perplexity's infrastructure, not on user-controlled hardware, unlike open-source alternatives.
- The feature improves task-specific performance, not general model intelligence.
| Feature | Perplexity Brain | OpenClaw + Mem0 | Hermes (Nous Research) |
|---|---|---|---|
| Hosting | Perplexity cloud | Self-hosted | Self-hosted |
| Memory type | Context graph + LLM wiki | Markdown + SQLite FTS5 | Extracted skill files |
| Data ownership | Perplexity-controlled | User-controlled | User-controlled |
| Access | Max/Enterprise Max ($200/mo) | Free, open-source | Free, open-source |
Brain remains in Research Preview, and Perplexity says further capabilities are planned without a specified timeline. For now, it's a clear signal that the next phase of the agent wars isn't just about smarter models — it's about which systems remember well enough to stop repeating their own mistakes.