Perplexity Details Brain's Architecture: Filesystem Memory, Dream Agents, and Production Results

Perplexity has published a detailed technical breakdown of Brain, the self-improving memory system for its Computer agent platform it first previewed on June 18, 2026 — this time laying out the underlying architecture and sharing 30-day production results.

Details

• Filesystem-based memory: Brain materializes memory as files in a memory/ directory with three layers — knowledge/ (linked Markdown pages synthesizing entities and learnings, the “Brain” itself), notes/ (topical snippets), and sessions/ (raw transcripts, summaries, indexes)
• Wiki-style linking: Pages use “wikilinks” ([[page]]) to connect related entities and “citation links” ([cite:N]) that trace claims back to the source session, so agents can verify where a piece of knowledge came from
• Agents browse it like a filesystem: Agents use ordinary operations — cat, grep, link-following, and Git-revision comparison — to explore Brain rather than querying a database
• Hybrid materialization: Instead of copying the entire memory tree (expensive) or reading it over a remote filesystem (Perplexity says 400-500x slower), relevant files preload at sandbox boot and a dedicated Memory Agent subagent pulls in more files on demand via semantic search
• Offline “Dream” agents: Background processes continuously refine Brain in a four-phase loop — orient to current state, summarize new sessions, attach notable observations to the right wiki pages, and update links/citations. Proposed changes go through deterministic format checks and semantic evidence verification before being committed via Git
• Offline eval results: Across 640 synthetic personas, answer correctness rose 6.1 points (60.0% → 66.1%) and evidence recall rose 5.2 points (57.3% → 62.5%), with the largest gains on preference questions (+10.2pp), temporal reasoning (+8.6pp), and activity-detail questions (+6.9pp)
• 30-day production results: Answer correctness +9.3 points, currentness +8.0 points, evidence recall +8.9 points, alongside roughly 15% fewer tokens used, 10% lower cost, and 10% faster generation

What happened next

This follow-up moves Brain from a research-preview announcement to a documented system with real production numbers, and the architecture — plain Markdown files, Git-backed edits, wiki-style linking — reads as a deliberately low-tech, auditable approach to agent memory rather than a black-box vector store. Perplexity says memory keeps improving with every session for users who have Brain enabled.