Memori is a persistent memory layer designed to provide AI agents with long-term recall and context-aware interaction capabilities. It functions as a middleware that automatically captures, structures, and stores agent execution traces and conversation history, allowing developers to inject relevant historical facts into model prompts to maintain continuity across sessions.
The system distinguishes itself through a dual-model storage approach that maintains information as both structured relational primitives for precise recall and rolling summaries for situational awareness. By utilizing middleware-based request interception, it captures interaction data without requiring manual changes to existing application logic. It further supports multi-tenant memory scoping, which isolates data using unique entity and process identifiers to ensure secure, context-aware retrieval in shared environments.
The platform covers a comprehensive lifecycle for memory management, including asynchronous background processing for embedding generation and knowledge graph construction to minimize latency. It provides a schema-agnostic database abstraction that connects to various relational and document-oriented storage backends, alongside observability tools for visualizing entity relationships and monitoring memory performance.
The project is implemented in Python and provides a command-line interface for managing environment settings, usage quotas, and database provisioning.