Perplexity Launches Brain

Justy Okay, this one actually matters. If an agent keeps making the same dumb mistake every day, and then one day it just… stops, that's a real product shift.

Cody Yeah. And Perplexity's angle is smarter than the usual memory pitch. It's not "remember my favorite formatting," it's "remember the dead end you hit in the repo and don't do that again."

Justy Which is such a cleaner user story, Cody. Nobody wakes up wanting a more intimate chatbot. They want the thing to stop re-learning the same project every session.

Justy Also, my week is apparently just me reading agent memory launches and muttering "okay, but does it keep receipts" like a tiny auditor. Nine months of this show has broken my brain before Perplexity launched Brain.

Cody That's fair. Mine is just a low-grade suspicion of any product called Brain. It sounds like the folder name for a startup deck, not a shipped feature.

Cody But this is shipped enough to talk about. Research Preview today for Perplexity Max and Enterprise Max, attached to Computer, and the core claim is a self-improving memory system built around the agent's work history.

Justy Right.

Cody It builds a context graph from sessions, connector results, document changes, and user corrections. Then at intervals, basically overnight, it synthesizes that into an L L M wiki that's automatically loaded into the sandbox for future runs.

Justy That auto-loaded wiki part is the adoption hook for me. If I'm already paying for Computer, I don't want to hand-curate some precious memory vault. I want the agent to show up tomorrow already knowing which project, people, files, and weird little corrections mattered.

Cody Mm-hm.

Justy And the target user is obvious. Repeated work with enough continuity that history pays off. Pipeline audits, support triage through connectors, debugging across repos… the examples are kind of textbook, but they're plausible.

Cody The technically interesting bit is the separation between logging and learning. They're not saying the model magically remembers. They're saying the system records traces, then runs a synthesis pass later that turns corrections into reusable lessons.

Justy Yeah.

Cody That's way more believable than "persistent memory" as vibes. It's basically a structured postmortem loop. The article's toy Python even makes that explicit with a synthesize step that promotes a correction into a lesson for the next day.

Justy And honestly that's the part I like most. This is very much our receipts thing again. Every memory entry links back to the session, file, or source it came from, which means when the agent gets weird you can inspect where the weirdness came from.

Cody Exactly.

Cody Without that, memory is just hidden state with a branding layer. With traceability, it's at least debuggable. This is basically their version of what we've been calling the boring controls winning, except wrapped in a much shinier box.

Justy Okay, but your skepticism is about to arrive on schedule like the overnight sync.

Cody I mean, yes. The numbers are first-party and early: plus twenty-five percent correctness on tasks it's seen before, plus sixteen percent recall, minus thirteen percent cost on tasks needing historical context. Those are decent, but they're also exactly the kind of metrics that can look great when the task shape already favors memory.

Justy I think that's fair. The thing I'm less worried about is whether memory helps repeated tasks. Of course it does. The real questions are whether the overnight cadence is fast enough, and whether teams are comfortable persisting this much work history in a graph in the first place.

Cody Oh interesting.

Cody Yeah, and the help-center bit says you can turn it off, which matters. But governance is still the adult question here. If Brain is pulling meeting notes, issues, product work, Slack threads, call transcripts, P Rs… cool, that's useful, and also that is a LOT of organizational memory to centralize.

Justy This is where I end up pretty positive, though. Not because I think Perplexity solved agent memory forever. Nobody has. But moving memory from "know me better" to "do the job better" is the right product reframe, and it's concrete enough that teams can actually decide if they want it.

Cody Yeah, no, I think you're right on that. It's also a nice antidote to the cat bureaucracy problem… not seven layers of mystical agent cognition, just a graph, a nightly synthesis pass, and better retrieval on the next run.

Justy Horrible phrase. Tragically useful. Okay, you've made me like this slightly more, which is annoying, so let's leave it there, Cody.