# MIGRO preprint reports smaller memory storage across language model handoffs

_Published Saturday, November 14, 2026 at 7:06 AM EST · AI, Science · Latest · Tier 2 — Notable_

Researchers presenting MIGRO report that their language model memory system scores 7.3% higher while using 99.3% less private memory than the strongest quality baseline on matched LoCoMo states. The preprint describes a persistent store that successive models can read and extend.

MIGRO preserves exact historical records, uses a shared semantic index to retrieve evidence, and converts selected records into inputs each receiving model can use. Reversible encoding allows exact recovery from compressed storage.

After a model handoff, continued writing retains 99.9% of full-history rebuilding quality. On million-token BEAM histories, private storage averages 57.5% less than the most compact baseline; the comparisons concern different baselines and workloads.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2611.15315)

---
Canonical: https://techandbusiness.org/newswire/UWiukoQDmN8m7fBM5KdG0K
Published: 2026-11-14T12:06:13.983Z
Story chronology: 2026-11-14T05:00:00.000Z
Retrieved: 2026-11-15T00:01:46.718Z
Publisher: Tech & Business (techandbusiness.org)