A Three-Layer Memory Architecture for LLMs
This article introduces Memento, a three-layer memory architecture for large language models (LLMs) that combines Redis, Postgres, and vector search to create an AI that can remember its experiences and learn from them over time.
Why it matters
This novel memory architecture could enable a new generation of AI assistants that can learn and improve over time, providing more personalized and effective support for users.
Key Points
- 1Memento builds an AI that knows the user, their project, and their mistakes, unlike traditional LLMs that forget everything after a session
- 2The three-layer architecture uses Redis for fast keyword indexing, Postgres for metadata storage, and pgvector for semantic search
- 3Memories have a temperature system that keeps frequently used memories hot, and automatically consolidates and forgets unused memories
- 4The system supports features like causal graph exploration, contradiction detection, and a closed learning loop to continuously improve the AI's performance
Details
Memento is a novel approach to building LLMs that can retain and learn from their experiences, unlike traditional models that forget everything after a session. The three-layer architecture combines Redis for fast keyword indexing, Postgres for metadata storage, and pgvector for semantic search. Memories are stored as atomic fragments (1-3 sentences) and have a temperature system that keeps frequently used memories hot, while automatically consolidating and forgetting unused memories over time. The system also supports features like causal graph exploration, contradiction detection, and a closed learning loop to continuously improve the AI's performance on the user's specific tasks and projects. This approach aims to create an AI assistant that genuinely gets better at working with the user over time, rather than just retrieving information from a static library.
No comments yet
Be the first to comment