What question did this study set out to answer?

The aim is to develop a universal memory architecture for AI that effectively manages episodic and semantic memory.

January 18, 2026Open Access

The Universal Memory Model (UMM): A Multimodal Architecture for Persistent Episodic–Semantic Memory in Artificial Intelligense Systems

SDSantosh K. DasariThe University of Texas MD Anderson Cancer Center

Key Points

The aim is to develop a universal memory architecture for AI that effectively manages episodic and semantic memory.
Developed a universal memory architecture called the Universal Memory Model (UMM)
Integrated various processes including canonicalization, embedding-based encoding, and attention-based retrieval
Addressed issues like context window limitations and memory drift
The UMM effectively supports persistent episodic and semantic memory
Demonstrated improvements in managing multimodal inputs
Enhanced long-term knowledge stability was observed

Abstract

This paper proposes a universal memory architecture for AI systems that supports persistent episodic and semantic memory across multimodal inputs. The Universal Memory Model (UMM) integrates canonicalization, embedding-based encoding, salience filtering, episodic storage, semantic consolidation, contradiction handling, and attention-based retrieval. The model addresses context window limitations, memory drift, and long-term knowledge stability.

KI fragen

Bookmark

View Full Paper