What matters in AI.

Subscribe

Amazon launches agent memory that cuts input text 12 times

Amazon tells that its memory system cut input from 167,000 to 13,500 tokens for each question and kept 93.0% correct answers.

Amazon Web Services made Amazon OpenSearch Memory available to all customers. It is a system that keeps the facts that an AI agent learns in one session. The agent can use these facts in a new session. Amazon tells that the system does the storage, the retrieval and the clean up of facts, and that teams do not operate servers.

How it works

The app writes each turn of a conversation to a space, which has the memory. The system keeps the first text and makes it easy to find in some seconds. It then makes short facts from the text and changes a fact when new data shows it is not correct. Before the agent answers, the app sends a query for the facts that fit.

The numbers

Amazon did the test on LongMemEval-S, a public test with 500 questions. Correct answers were 93.0%, against 95.2% when the agent read the full history. The result was 4.4 points higher than a keyword query over raw text. Amazon did the test one time. The range is 2 to 3 points more or less.

What is not known

The result is for one set of models and adjustments. Amazon tells that other models and adjustments can give other results. No external team did a test. The source gives no price.

Insight: Cloud vendors sell agent memory

In October 2026, Databricks made agent memory available in Beta and told of 31% less token use. Amazon tells of 12 times less input text. Each memory stays in the storage of the vendor that sells it. This reads as a race to own where agent knowledge stays, because a team can find it hard to go. Check how you can export it.

Sources

  1. Introducing Amazon OpenSearch Memory: Persistent memory for AI agentsaws.amazon.com
AI MATTER · NEWS · AI MATTER · NEWS ·18 NOV2026

Posted

Tags

Learn the terms in this story

More in Agents

All Agents news