Perplexity’s Brain gives AI agents a memory they can inspect and edit

Perplexity has set out how Brain stores an AI assistant’s past work and corrections outside the model, then brings relevant information into later tasks. The practical point: users can inspect, edit or delete that memory, while developers using Perplexity’s Agent API must build their own persistent memory system.

Perplexity Watch analysis

What happened

Perplexity’s Brain, part of Perplexity Computer, organises past sessions into summaries, subject notes and linked wiki pages. Background agents called Dream process session records overnight and can revise the wiki; Perplexity says staged changes are checked before they are committed, with a Git diff available to inspect. Memory entries link back to their sources, and users can edit or delete them in a Memory view. Perplexity says Brain excludes incognito sessions and has a separate toggle.

In a guide published on 2 November, Unite.AI also describes how developers can carry conversation context between Agent API requests, use versioned Skills and Profiles, and connect agents to developer-controlled storage. Perplexity distinguishes that toolkit from Brain: the Agent API does not come with a self-maintaining memory system. Read the guide.

Key findings

• Brain keeps memory outside the model’s weights
  Changing a wiki page changes what the agent can retrieve later, without retraining the model.
• Developers retain responsibility for persistent memory
  The Agent API offers ways to pass context and use external stores, but developers decide what to save, retrieve and remove.
• Perplexity reports gains in an internal evaluation
  The company says Brain-enabled sessions improved correctness by 9.3 percentage points and recall by 8.9, while using about 15 per cent fewer tokens in a paired 30-day evaluation.
• The company’s example leaves attribution murky
  Perplexity says a customer workflow saved 120 work hours, but could not quantify how much of that came from memory rather than other tools and automation.

Why it matters

Persistent memory can make an agent more useful, but it also raises ordinary, important questions: what does it retain, where did a remembered instruction come from, and can a user correct or remove it? Perplexity’s described controls and source links give users something more tangible than the familiar promise that an agent will simply remember better.

For developers, the distinction between Brain and the Agent API matters before building. The API can carry conversation state, but episodic records and durable knowledge remain the application’s job. That is more control, and more plumbing.

Our read

The most useful part is not the word “memory”; it is the trail back to sources and the ability to inspect, edit or delete what has been stored. Perplexity’s reported benchmark improvements are encouraging, but they are company-reported results, not a general guarantee that memory improves every agent. Developers should test whether an agent uses verified, current information and respects corrections, rather than treating recall as the whole scorecard.

What to watch

• Whether Perplexity publishes more detail about the Brain evaluation and its test set.
• How users can manage memory in practice, including corrections and deletion.
• Whether developers using the Agent API adopt clear ways to track sources and stale information.

Discussion spark: Should an AI assistant’s persistent memory be managed mainly by the product provider, or should users and developers control the store themselves?

not affiliated with or endorsed by Perplexity