A Model-Agnostic Persistent Memory Infrastructure for Long-Term AI Agents Using Hybrid Retrieval, Scope Isolation, and MCP

This paper presents a model-agnostic persistent memory infrastructure for AI agents, designed to support long-term information storage and retrieval across sessions, models, and agent frameworks. The proposed architecture explores hybrid retrieval combining semantic search and BM25, Elasticsearch-based memory storage, REST API and Model Context Protocol (MCP) interfaces, and mechanisms for memory lifecycle management, deduplication, conflict handling, and scope isolation. The work focuses on the architectural design and implementation of reusable memory infrastructure for AI agents. Empirical performance claims should be interpreted in accordance with the validation methods and experimental evidence reported in the paper.