Rosetta Memory: Adaptive Memory for Cross-LLM Agents

2026-06-05 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Rosetta Memory is a novel adaptive memory system designed for cross-LLM agents, addressing the challenge of enabling memory written by one Large Language Model to be effectively consumed by another. Unlike traditional LLM-centric memory designs, Rosetta Memory adopts a memory-centric approach to facilitate adaptation when users frequently switch between models like Claude for coding and GPT for writing, or route different task steps to various backbones for cost efficiency. The system tackles upstream-downstream memory adaptation through both write and read operations, employing two jointly trained, profile-conditioned operators that optimize how memory is stored and presented for enhanced task completion. To ensure broad generalization across diverse LLMs, it incorporates a minimum-gain sampling curriculum, prioritizing less-served models during training. A performance-gap reward function measures the operators' true contribution against a naive baseline. Experiments on datasets such as HotpotQA, 2WikiMultihopQA, and MuSiQue demonstrate consistent outperformance against baselines and robustness to unseen model replacements.

Key takeaway

For Machine Learning Engineers designing or managing multi-LLM agent systems, Rosetta Memory offers a critical solution for memory interoperability. You should consider integrating memory-centric adaptation to ensure seamless knowledge transfer when switching between models like Claude and GPT, or routing subtasks to different LLMs. This approach enhances agent persistence, improves long-horizon planning, and allows for more cost-effective model utilization without sacrificing performance, even with unseen models.

Key insights

Rosetta Memory enables cross-LLM agents to share and adapt memory effectively, shifting from LLM-centric to memory-centric design.

Principles

• Memory adaptation is critical for cross-LLM agent persistence.
• Prioritize least-served LLMs for broad generalization.
• Measure true operator contribution via performance-gap reward.

Method

Rosetta Memory jointly trains two profile-conditioned operators for memory storage and presentation, optimized for task completion. It uses a minimum-gain sampling curriculum and a performance-gap reward for training.

In practice

• Facilitate seamless LLM switching in agent workflows.
• Route tasks to different LLMs for cost-effective trade-offs.
• Improve long-horizon planning with adaptive memory.

Topics

• LLM Agents
• Adaptive Memory
• Cross-LLM Systems
• Memory Management
• Multi-model AI
• Agent Persistence

Best for: Research Scientist, AI Architect, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• AMA: Adaptive Memory via Multi-Agent Collaboration — AMA uses multi-agent collaboration and hierarchical memory to provide adaptive, consistent long-term memory for LLM agents.
• Enhancing Software Engineering Through Closed-Loop Memory Optimization — MemOp grounds memory utility in validated downstream impact to optimize LLM-based SE agents.
• Phase Transitions in Agent Memory: Recurrent Memory — Long-running LLM agents benefit from density-driven memory consolidation, triggering processing only upon sustained semantic recurrence.
• From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms — LLM agent memory evolves through stages of storage, reflection, and experience, driven by consistency, dynamic environments, and continual learning.
• Engram: How LLMs Finally Get Scalable Memory — Engram enhances LLM factual recall and reasoning by integrating a scalable, hash-based memory lookup system for direct knowledge retrieval.
• MeMo's memory model lets teams upgrade their LLM without retraining it — and performance jumps 26% — MeMo uses a separate, smaller memory model to update LLM knowledge without retraining the main reasoning engine.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.