Perplexity 详解 Brain 记忆架构与 Agent API 测试方法

Perplexity on November 2, 2026, published a technical guide to persistent AI agent memory, detailing the Brain system in Perplexity Computer, the Agent API’s memory-building components, and tests for whether stored context improves an agent’s later work.

Brain’s Architecture and Reported Results

Perplexity describes Brain as its self-improving memory system for Computer, built to organize past work and corrections into an evolving knowledge wiki. It uses token-level memory, meaning information is stored outside the model and read into context when needed. Changing a wiki page alters what the agent can read on a later task without modifying the model’s learned weights.

Brain stores these records as files in the agent’s sandbox across three layers. The sessions/ directory retains raw transcripts and their summaries, notes/ organizes distilled information by subject, and knowledge/ contains the wiki’s linked Markdown pages about people, projects and concepts. The guide classifies session records as episodic memory and the wiki as semantic memory; Skills and saved instructions provide procedural memory for how tasks should be performed.

Background agents called Dream process each day’s session records overnight, produce summaries and revise the wiki when information has changed. Perplexity says it checks staged edits before they are committed through Git, leaving a diff available for inspection. Each memory entry links back to its source, while Brain’s background updates reinforce information that remains valid, revise entries after changes and mark outdated material.

Perplexity reported that Brain-enabled sessions in a 30-day paired evaluation improved correctness by 9.3 percentage points and recall by 8.9, while using roughly 15% fewer tokens. In a separate internal benchmark covering 150 financial-document questions, an exploration agent using standard file tools reached an answer score of 0.627, compared with about 0.37 for semantic search over the same corpus.

Agent API Memory Components

Perplexity explicitly distinguishes its developer offering from Brain: the Agent API does not include a self-maintaining memory system. It instead provides versioned procedures, methods for carrying conversation context between requests, and tools for reading from and writing to a developer-controlled store.

A custom Skill bundles instructions, procedures and supporting files under a version number. A Profile saves the broader configuration, including the model, instructions, tools, Skills, connectors and run settings, under one versioned ID. Skills do not learn independently; their instructions change when a person or automated process uploads a new version.

Developers have two options for preserving working context between turns. They can place selected earlier messages in the next request’s input array, deciding what to retain, summarize, redact or drop. Alternatively, they can pass a completed response’s ID as previous_response_id, allowing Perplexity to maintain the conversation state on its servers without the application resending earlier turns. Either approach still requires any new instructions, tools or model settings needed for the next response.

Episodic and semantic storage remain the developer’s responsibility. Function calling can implement a read operation before the agent acts and a write operation afterward, with the application executing calls and returning results. Recording the outcome, supporting source and subsequent user correction provides the beginning of an episodic record rather than merely logging that information was retrieved. A compatible remote MCP server offers another route by exposing memory operations without separate custom-function wrappers. Perplexity lists GitHub, Slack, Google Drive, Datadog, Linear and Notion among its managed connectors, while custom connectors can provide access to internal systems.

Testing and Computer Controls

The guide’s fresh-session test starts a request without replayed messages or previous_response_id, while preserving access to the same memory store. The agent is then asked to perform related work without receiving the earlier correction again. For a deployment agent, Perplexity proposes checking that the agent retrieves and uses the verified rollback command for the affected service, leaves other services’ procedures unchanged, and adopts newer guidance when the command changes while retaining a traceable source.

For comparisons with and without memory, Perplexity recommends holding the model, Profile and Skill versions fixed and starting from a known store snapshot so one test cannot teach the agent the answer to the next. It also recommends saving failed cases for subsequent revisions to pass. Rather than relying on recall alone, the guide says to measure whether memory-dependent tasks are completed, with cost and latency included.

In Computer, Brain draws on sessions, connected tools, files and artifacts, and user corrections. Its Memory view lets users inspect entries, follow them back to their sources, and edit or delete them. Brain excludes incognito sessions and has a separate toggle.

The guide’s Rho example describes an invoicing project in which Computer reviewed changes overnight, flagged inconsistencies and published a morning digest in Slack. Rho reported reducing weekly meeting time from 2.5 hours to 15 minutes and estimated saving 120 work hours over the 12-week project. Perplexity said the 120-hour estimate covered the combined workflow of connectors, overnight runs and the Slack digest, and that it could not yet quantify memory’s separate contribution.

The guide therefore leaves memory’s individual contribution to Rho’s reported savings unresolved.