MIGRO: A Decoupled and Portable Long-Term Memory System for Language Models
Summary
- Researchers introduced MIGRO, a decoupled long-term memory system designed to act as a persistent state shared across successive language models.
- MIGRO uses asymmetric fidelity with deferred selectivity, preserving historical evidence exactly while organizing it approximately for query-time retrieval.
- Host-specific interfaces convert selected evidence to accommodate different model vocabularies and tokenizations, while reversible encoding reduces storage costs.
- On matched LoCoMo states, MIGRO scores 7.3% higher and uses 99.3% less private memory than the strongest quality baseline.
Why it matters
MIGRO provides a scalable memory architecture that allows accumulated interaction history to persist and grow across model upgrades without requiring full-history rebuilding.