Understand this story
On November 2, 2026, Perplexity released a technical guide introducing the architecture of the Brain memory system in Perplexity Computer, the memory-building components of the Agent API, and testing methods for verifying whether stored context can improve subsequent tasks. The guide explains that Brain stores conversation records, notes, and a knowledge wiki as files in a layered manner within the agent sandbox; a background agent, Dream, processes each day's conversations every night and revises the wiki, with changes committed after being checked via Git staging. Perplexity says that in a 30-day paired evaluation, sessions with Brain enabled saw a 9.3 percentage point increase in accuracy and an 8.9 increase in recall, with token usage reduced by about 15%; the Agent API itself does not include a self-maintaining memory system, and developers must manage episodic and semantic storage themselves.
AI-generated from public reports
Latest development · Report
Perplexity Details Brain Memory Architecture and Agent API Testing MethodsDevelopments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 3
Unite.AI
Perplexity Details Brain Memory Architecture and Agent API Testing Methods
On November 2, 2026, Perplexity released a technical guide introducing the architecture of the Brain memory system in Perplexity Computer, the memory-building components of the Agent API, and testing methods for verifying whether stored context can improve subsequent tasks. Brain stores conversation records, notes, and a knowledge wiki as layered files in the agent sandbox; a background agent, Dream, processes each day's conversations every night and revises the wiki, with changes committed after a Git staging check. Perplexity says that in a 30-day paired evaluation, sessions with Brain enabled saw a 9.3 percentage point increase in accuracy and an 8.9 increase in recall, with token usage reduced by about 15%; the Agent API itself does not include a self-maintaining memory system, and developers must manage episodic and semantic storage themselves.