Reverie
Persistent AI memory achieving 94.6% on LongMemEval
A layered cortex architecture that gives AI systems long-term memory — approaching human-like recall across sessions, topics, and time.
Read the PaperLongMemEval Leaderboard
Top systems ranked by overall accuracy on the LongMemEval benchmark (n=500, GPT-4o judge).
| Rank | System | LLM | Score |
|---|---|---|---|
| 1 | Mastra OM | GPT-5-mini | 94.87% |
| 2 | Reverie | Sonnet | 94.6% |
| 3 | Hindsight | Gemini 3 Pro | 91.4% |
| — | Oracle baseline | GPT-4o | 82.4% |
How It Works
Reverie organizes memory into five cortical layers — from raw episodes to abstract patterns to persistent identity — with bidirectional signal flow that mirrors biological memory consolidation. No vector-database-and-pray. Structure earns recall.