# Google reports model memory consolidation across 10M-token benchmark contexts

_Published Tuesday, October 27, 2026 at 2:02 PM EDT · AI, Science · Latest · Tier 2 — Notable_

![Google reports model memory consolidation across 10M-token benchmark contexts — Primary](https://storage.googleapis.com/gweb-research2023-media/images/HO_previewImage1.width-800.format-jpeg.jpg)

Google Research says its experimental "sleep" framework retained information from contexts up to 10M tokens on the BABILong benchmark by consolidating it into model parameters. The approach alternates active use with offline training to incorporate new knowledge while limiting the loss of previously learned capabilities.

During offline training, the model activates pre-allocated memory blocks and updates numerical weights using self-generated practice examples. Google also reported near-zero performance decline in sequential translation evaluations involving Manchu and Kalamang. The framework operates within a fixed architecture and human-defined training rules; its reported gains come from benchmark evaluations.

## Sources

- [The latest research from Google](https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/)

---
Canonical: https://techandbusiness.org/newswire/bwytQDaO4jrmepuWrA6N7s
Published: 2026-10-27T18:02:32.429Z
Story chronology: 2026-10-27T17:35:28.000Z
Retrieved: 2026-10-27T20:39:31.738Z
Publisher: Tech & Business (techandbusiness.org)