# Preprint demonstrates shared internal memory for language model agents

_Published Friday, October 16, 2026 at 5:01 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers report a way for language model agents to share internal computational state through a common pool without fitting a separate transfer mechanism for each pair of agents. In their preprint, a small trained adapter lets a receiving model consult another model's stored attention data while retaining its own, matching the accuracy obtained by reading the source document.

The same adapter also works with previously unseen publishers and several publishers together, provided their internal states resemble the receiver's. Directly inserting intermediate activations works better than presenting them alongside the receiver's own, but requires knowing where those activations belong.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2610.22026)

---
Canonical: https://techandbusiness.org/newswire/7hM8guhwgqkR7pSQI4aIGl
Published: 2026-10-16T09:01:18.745Z
Story chronology: 2026-10-16T04:00:00.000Z
Retrieved: 2026-10-16T11:34:45.238Z
Publisher: Tech & Business (techandbusiness.org)