Perplexity Brain memory: how stored context reaches new tasks

Perplexity’s November 2, 2026, technical guide to agent memory describes Brain as a persistent-memory system for Perplexity Computer. It organizes work history into linked records and retrieves relevant context for later tasks. The Perplexity Agent API offers ways to carry conversation context forward, but developers must build and operate persistent memory themselves.

The practical distinction is retrieval: information can remain stored without being surfaced when a new task needs it. For a separate look at where Computer runs on a Mac and which actions still use connected services, see NeoTeo’s earlier article on Perplexity Computer’s local mode.

Brain organizes work history into linked memories

Perplexity describes three file layers: sessions/ holds raw transcripts and summaries, notes/ contains distilled information organized by topic, and knowledge/ contains linked Markdown wiki pages. The guide maps sessions to episodic memory, or what happened; the wiki to semantic memory, or what the agent treats as true; and Skills and instructions to procedural memory, or how to perform a task.

Brain stores these records outside the model’s learned weights and loads relevant content into its context when needed. Editing a wiki page changes the stored information, not the model’s learned parameters.

Perplexity says background agents called Dream process daily sessions overnight, summarize them and propose wiki changes. The changes are staged and checked through Git, so they can be inspected. That gives Brain a maintenance loop: past work can be distilled into records that are revised as new sessions and corrections arrive.

How saved context reaches a later task

Perplexity’s description separates persistence, which keeps information available, from retrieval, which brings relevant information into a new task. An index points Computer to a memory page; Computer can read that page, search files, and follow wiki links or citations back to the original conversation or source. If the relevant record is not found and loaded, its presence in storage alone does not supply it to the task.

The Memory view presents linked items as a graph alongside categories such as notes, concepts, entities and workstreams. The graph is an interface for viewing memory, while Perplexity describes the retrieval path separately.

This difference matters when you evaluate an agent: a fact that was saved in an earlier session is not, by itself, evidence that the agent will use it later. Perplexity recommends a fresh-session test in which the task begins without the earlier messages or a conversation-continuity ID, while retaining access to the same memory store.

Keeping records useful as facts change

Perplexity identifies four common memory-system problems: records can grow without bound, old beliefs can become outdated, limited capacity can force forgetting, and retrieval can fail to improve from later outcomes. Its suggested maintenance cycle is to ingest, revise, forget and retrieve. For changing facts, the guide recommends recording when a claim became valid, its source and what it replaced; it also advises keeping unresolved source disagreements visible.

The guide summarizes a selective-memory method from Workday AI Research that screens candidate memories for usefulness, trustworthiness, novelty, freshness and whether a fact is durable or temporary. Perplexity attributes to that method 12% higher precision in keep-or-drop decisions while retaining 97% of relevant memories. It also reports that omitting the durable-versus-temporary distinction reduced decision quality by roughly 18%. These figures are Perplexity’s summary of the Workday approach.

For developers, the practical lesson is to preserve provenance and time context, not just a short answer. A current setting and a temporary condition should not become indistinguishable from a durable project decision.

Brain and the Agent API handle memory differently

Brain is a managed memory loop inside Perplexity Computer. The Agent API gives developers continuity tools and ways to connect their own memory operations, but Perplexity says it does not include a self-maintaining Brain-like system.

For an API project, conversation continuity and persistent memory solve different jobs. Replaying messages or passing previous_response_id carries conversation state; it does not replace a store and update logic that preserve useful facts across separate tasks. Developers who want persistent memory must design those operations and decide how records are revised or discarded.

What the reported tests measure, and what users can control

Perplexity reports that its 30-day paired evaluation found Brain-enabled sessions improved correctness by 9.3 percentage points and recall by 8.9 percentage points, while using roughly 15% fewer tokens. A separate internal comparison covered 150 financial-document questions: an agent exploring files with standard tools scored 0.627, while semantic search scored about 0.37. That second comparison measured file exploration against semantic search, not Brain against a version without memory.

To test memory in an application, Perplexity recommends holding model, Profile and Skill versions fixed, starting from a known memory-store snapshot, and testing in fresh sessions. It advises comparing task completion, cost and latency rather than recall alone, and saving failures as regression cases. A controlled snapshot helps prevent one trial from changing what a later trial can retrieve.

Perplexity’s Rho invoicing example reports that weekly meeting time fell from 2.5 hours to 15 minutes. The company estimates 120 work hours saved over the 12-week project, attributing that estimate to a combined workflow of connectors, overnight runs and a Slack digest. It says Brain’s separate contribution was not quantified.

Computer’s Memory settings include controls for search history, Notes, memory management and Brain. Perplexity says users can inspect entries and their sources, edit or delete memories, exclude incognito sessions, and use a separate Brain toggle.