Breaking

Perplexity Launches Brain, a Self-Improving Memory for AI Agents

June 19, 2026 at 15:10 EDT

AI Agents
Foundation Models

On June 18, 2026, Perplexity launched Brain, a self-improving memory system that lets AI agents learn from the tasks they perform, releasing it as a research preview for Perplexity Max and Enterprise Max subscribers.

June 18, 2026 · Perplexity

"Brain": an AI agent that learns from its own work

Perplexity's new self-improving memory system records every task into a context graph and synthesizes it overnight — so the agent gets better the more it repeats similar work.

+25%

Answer correctness on repeated tasks

+16%

Recall on repeated tasks

−13%

Cost for tasks needing historical context

Early measured impact on repeated tasks

Bars drawn to scale — positive gains (green/blue) vs. cost reduction (orange).

+25%

Answer
correctness

+16%

Recall

−13%

Token cost
(historical)

How the self-improving loop works

1 · Do the work

Records successes, failures, corrections & sources

→

2 · Build the graph

Each task added to a traceable context graph

→

3 · Synthesize overnight

Updates a user-specific "LLM wiki"

→

4 · Carry context forward

Loaded into the next task's sandbox

The promise

A "work memory" that learns the user's environment — potentially competitive for ongoing research and agentic work, treating tokens as an investment in future efficiency.

The concern

Token consumption is heavy and efficiency is a question; some doubt it differs much from existing Spaces. Long-term feedback is only starting to accumulate.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.