AI Wednesday 07 : TITANS — Mamba, Titans, and the MIRAS Framework

By Sebin Thomas · April 29, 2026

Overview

This week's AI Wednesday looked at a new wave of sequence model architectures beyond the standard transformer. We explored Mamba's selective state space design, Google's Titans architecture for long-term memory, and the MIRAS framework that unifies how modern models handle memory, attention, retention, and online learning.

Topics

• Mamba: selective state space models and linear-time sequence processing
• Titans: neural long-term memory with test-time memorization
• MIRAS: a framework for designing memory-driven sequence models
• How these architectures compare to transformers on long-context tasks
• The broader shift toward efficient, memory-aware sequence modeling

Resources

• Mamba: Linear-Time Sequence Modeling with Selective State Spaces
• Titans + MIRAS: Helping AI have long-term memory (Google Research blog)
• Titans: Learning to Memorize at Test Time
• MIRAS: It's All Connected — A Journey Through Test-Time Memorization, Attentional Bias, Retention, and Online Optimization

Photos

Highlights

• MIRAS provides a useful lens for understanding why so many recent architectures — from Mamba to Titans — are converging on associative memory as the core design problem.

Next Week

• Topic: LiteRT — Google's Edge AI Runtime
• Host: Sebin Thomas