On 2 November 2026, Perplexity published a technical guide to persistent agent memory that makes a distinction the market has blurred. Recall is not the test. The test is whether an agent does better work in a fresh session because of what it kept, revised, and can trace to a source. Perplexity's answer is Brain, the self-improving memory system inside its Computer product, which organises past work and corrections into an evolving knowledge wiki.

How persistent agent memory is built in Brain

Brain stores memory as files in the agent sandbox, not as weights in the model. Changing a page changes what the agent can read on the next task. The guide divides the store into three layers. The sessions directory holds raw transcripts and their summaries. The notes directory holds distilled information organised by topic. The knowledge directory holds the wiki itself, linked Markdown pages about people, projects, and concepts.

Perplexity maps those layers to memory types drawn from the CoALA paper. Sessions are episodic memory, what happened. The wiki is semantic memory, what the agent takes as true, with links back to evidence. Skills and saved instructions are procedural memory, how to do things. The conversion matters more than the storage. A session records that a user rejected a PNG file. The instruction to use PDFs becomes a preference the agent carries forward. A correction to the export process becomes a procedure it can follow next time, in the guide's framing, from this happened, to this is true, to this is how we do it.

Background agents called Dream do the conversion overnight. They read the day's sessions, summarise them, and revise the wiki where the picture has changed. Perplexity's deeper technical account adds the control path. Dream writes proposed changes into a staged output tree. Deterministic validation checks that pages are well formed, including required frontmatter and citation format. Semantic verification checks that a proposed synthesis is supported by the gathered evidence and consistent with the rest of the graph. Only then does a controlled sync commit the diff through Git, where the version history stays inspectable. Each wiki claim carries citation edges down to the raw session or connector source that supports it, and wikilinks connect pages laterally to related people, projects, and concepts.

What the numbers show, and what they do not

The headline result is from a 30-day paired evaluation. Brain-enabled sessions improved correctness by 9.3 percentage points and recall by 8.9 points, while using roughly 15 percent fewer tokens. The technical account adds currentness up 8.0 points, cost down about 10 percent, and generation about 10 percent faster.

The offline numbers are smaller, and Perplexity publishes them in the same place. On an internal set of 640 questions across 44 synthetic personas, with each answer mechanically tied to evidence in the account history, the wiki raised correctness from 0.600 to 0.661, a gain of 6.1 percentage points, and evidence recall from 0.573 to 0.625, a gain of 5.2 points. The largest gains were on preference recall, up 10.2 points, and temporal reasoning, up 8.6 points.

The public benchmark ablations are mixed, by Perplexity's own account. Removing the wiki on LoCoMo reduced answer correctness by 4.6 percentage points on average over three runs. On LongMemEval-S it produced no statistically significant change, because that benchmark mainly tests recovery of facts from individual sessions, where the transcript already provides a path to the answer. Perplexity states that these experiments used benchmark subsets and are not definitive benchmark results. In a separate internal benchmark of 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.

The vendor caveat applies to all of it. These are Perplexity evaluations, on synthetic personas and subsets, with no independent replication reported in either post. The Rho example in the guide carries the same limit in miniature. A team using Computer on an invoicing project reported cutting weekly meeting time from 2.5 hours to 15 minutes and estimated saving 120 work hours over a 12-week project. Perplexity states that the estimate covered connectors, overnight runs, and a Slack digest together, and that it cannot yet quantify how much came from memory alone.

The developer product is parts, not a persistent agent memory system

The guide is explicit on a point buyers should not miss. Perplexity's Agent API does not ship a self-maintaining memory system like Brain. It provides pieces a developer assembles. A custom Skill packages instructions, procedures, and supporting files under a version number. A Profile saves the model, instructions, tools, Skills, connectors, and run settings under one versioned ID. Skills do not learn on their own; they change when a person or an automated process uploads a new version.

Working context can be carried in two ways. A developer can replay selected earlier messages in the next request and decide what to keep, summarise, redact, or drop. Or the developer can pass a completed response ID as previous_response_id and let Perplexity hold the conversation state on its servers. Episodic and semantic storage stay the developer's responsibility, through function calling with a read before the agent acts and a write afterwards, or through a compatible remote MCP server. Managed connectors named in the guide include GitHub, Slack, Google Drive, Datadog, Linear, and Notion.

The Revision Ledger: a test a company record can pass

Perplexity's own test is the most portable part of the guide. Close the conversation. Start a new request without replaying messages and without previous_response_id, but give the agent access to the same memory store. Then check three things. The correction survives: the agent retrieves and uses the verified rollback command for the affected service. The correction stays in scope: it leaves other services' procedures alone. New information replaces old guidance: when the command changes, the agent uses the current version and can trace the change to its source. The guide adds the method discipline. Hold the model, Profile, and Skill versions fixed. Start each comparison from a known snapshot of the store, so one test cannot teach the agent the answer to the next. Save failed cases, and require the next change to pass them. Measure task completion, cost, and latency, not recall alone.

For a company record, that test needs one extension, which this article calls the Revision Ledger. Brain already stages its overnight edits, validates them, and commits them as an inspectable Git diff. A company needs the same ledger at record level. Every persistent entry should carry its source, its date, the version it superseded, and the diff that the overnight revision proposed, with a person able to inspect, edit, or delete the entry before it governs the next task. Perplexity gives Computer users a Memory view to inspect entries, follow them to sources, and edit or delete them, and it excludes incognito sessions. A shared company record needs that inspection right for the team that owns the record, not only for the individual whose session created it. That is the same source-first standard the recent memory research points to: raw retrieval beat dedicated memory systems in most head-to-head tests when the extracted store lost information the source still held.

Persistent agent memory earns its cost only when a correction made once changes work later, in a new session, with the evidence still attached. A wiki that rewrites itself overnight is a strong design for that job, and Perplexity's mixed ablations are an honest map of where synthesis helps, across sessions, and where a transcript or a search over the source is already enough. The question for any memory system, personal or company-wide, is not how much it recalls. It is whether its revisions can be inspected, reversed, and trusted by the next person, or the next agent, who acts on them. How that record is built, reviewed, and kept current is the subject of How Jilu Works.