Perplexity Brain Adds Self-Improving Work Memory to Its Agent

Perplexity's new Brain feature for its Computer agent addresses something that has quietly been a ceiling for anyone using AI agents in repetitive workflows: agents don't get better with use. Every run starts from scratch, repeating the same wrong turns, making the same tool calls that failed last Tuesday. Brain is Perplexity's attempt to fix that.

According to MarkTechPost's coverage, Brain builds what the company calls a context graph of completed agent work, logging actions, sessions, and corrections. Overnight, a synthesis process converts those logs into reusable lessons, stored in an LLM wiki that automatically loads into the agent sandbox on each new run. The company's early first-party testing reportedly shows +25% answer correctness on repeated tasks, +16% recall, and a -13% cost reduction on workflows that depend on historical context.

The framing Perplexity uses is worth pausing on. Traditional AI memory, the kind that remembers your name and preferred formatting, is about the user. Brain is about the agent's work: what succeeded, what failed, what required correction. That distinction reframes memory as a performance mechanism rather than a personalization feature, and it's a more honest account of what enterprise teams actually need from an autonomous agent.

The honest caveat is that all three headline numbers are first-party, announced at launch, with no independent benchmark validation published. Overnight batch updates also mean that errors from today's run don't feed back until tomorrow's synthesis cycle, which could matter in fast-moving workflows. What the reporting doesn't give you is any clarity on data governance: what exactly persists in the work-history wiki, who owns it in an Enterprise Max deployment, and how teams remove records when a project ends.

Brain is rolling out in Research Preview to Perplexity Max and Enterprise Max subscribers. If the correctness and cost numbers hold up under real workloads, data scientists running weekly pipeline audits and support teams routing tickets stand to benefit most. The broader signal worth watching is whether work-focused memory, rather than user-preference memory, becomes the standard approach as agentic tools mature.