Perplexity’s AI Agent Now Has a Brain That Learns From Its Own Mistakes

Perplexity launched Brain today, a memory system for its Computer agent that gets smarter the more you use it. Not by remembering your name or job title—by logging what the agent actually did.

Advertisement

“With Brain, Computer starts each task with full context of your projects, decisions, and sources instead of from scratch,” Perplexity says. “Every memory links back to the session, file, or source it came from with full transparency and control.”

Every time Computer completes a task, Brain adds it to a context graph. That graph tracks which connectors were used, which sources held up, what corrections the user made, and what didn’t work. 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 starts. Each memory entry links back to the session or file it came from, so you can trace any decision to its origin.

The logic is straightforward. Most AI memory is about you, the user—your preferences, your habits, your name. Brain’s memory is about the work. What the agent tried, what got corrected, what source led somewhere useful. That’s a more actionable kind of memory for any system meant to actually get things done.

Perplexity’s own early metrics show Brain boosts answer correctness by 25% on tasks Computer has already handled, improves recall by 16%, and cuts the cost of context-heavy tasks by 13%. Those are internal numbers, not third-party benchmarks. But the direction makes sense: An agent that starts each morning knowing which sources failed last week will waste fewer tokens finding that out again.

Not exactly new territory

Some may think this is pretty cool, others may be asking “am I missing something?” In reality, Perplexity is bringing a niche implementation into a mainstream audience.

OpenClaw—which has accumulated over 379,000 GitHub stars since launch—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. With the Mem0 plugin, memory capture happens automatically at the system layer and survives restarts and context compaction.

OpenClaw also 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.

Hermes, Nous Research’s self-improving agent released in February 2026, goes further. After each completed task, Hermes evaluates the outcome, extracts reusable reasoning patterns, and writes them as skill files in plain markdown. Next time it hits a similar problem, it loads the skill instead of reasoning from scratch.

It also has skills with similar propositions (like Obsidian Mind) that aim at making the agent more personal and more useful.

Both tools are self-hosted. You run them on your own hardware, your data stays there, and you control everything. That’s the fundamental difference from Brain. Brain is supported by a multibillion company, and runs entirely on Perplexity’s ecosystem.

Who this is for—and who it isn’t

Brain is for people already paying $200 a month for Perplexity Computer. If you’re running it for recurring work—competitive monitoring, weekly reporting, research tasks that reference prior runs—the upgrade is real. The agent stops reinventing the wheel every session.

That said, Brain is not a local memory tool you control. Perplexity’s infrastructure holds the context graph, the LLM wiki, and all session history. You get transparency into what’s stored, but not ownership of it. Users who need full data sovereignty are better served by Hermes or OpenClaw with plugins and skills like Mem0, Honcho, Obsidian Mind or Hindsight—which keep your data on hardware you own.

It’s also worth being clear about what “self-improving” does and doesn’t mean here. Brain makes Computer better at tasks it’s already done for you. It doesn’t make the underlying models smarter. Cross-domain generalization—taking what it learned helping with financial research and applying it to a coding task—remains an open problem Brain doesn’t claim to solve.

Brain is in Research Preview starting today for Max and Enterprise Max subscribers. Perplexity says new capabilities are coming soon, with no timeline attached.