# Anthropic Tests Find Claude Memory Gains on Some Document Tasks By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/51800 Published: 2026-11-08T20:39:40.000Z Updated: 2026-11-08T20:39:40.000Z Section: Technology > Claude’s score reached 48% across 25 questions after 128 sessions, while one legal benchmark showed no clear benefit. Claude agents scored higher on some document tasks after organizing information from earlier sessions, but the gains varied by test and one memory wiki grew larger than its source files. In a document test, Claude faced 25 questions about an archive of roughly 350,000 files containing about 420 million tokens. The wiki-only score rose from 13% of graded criteria after two sessions to 48% as the wiki was built over 128 sessions. Pairing the wiki with search lifted the score from 54% to 61% as the memory base grew; search alone scored 53%. The combined setup also saved about 20,000 tokens per question. On a separate set of 24 merger questions, Claude scored 87% with a wiki, compared with 73% using search alone. The results differed across tasks. A wiki showed no clear benefit on one legal benchmark, and another example grew larger than the source documents it summarized. The findings point to gains in some settings, alongside cases where an expanded memory store added little or became unwieldy. Claude’s agent memory holds information for use in later sessions. Dreaming reviews that stored memory and past session transcripts, then creates a separate, reorganized store. The process updates retained information rather than Claude’s model weights. The original store stays intact, and developers can review or discard the generated version. They can also choose automatic updates or review proposed changes first. The feature is part of [Claude Managed Agents, which entered public beta April 8, 2026](). Dreaming remains in research preview and is available by request. Lance Martin, author of the post, wrote, “Claude can make incorrect predictions or write rules to memory that are wrong from the start.” He added, “Dreaming is one way to fix these mistakes, but it requires compute.” ## Links in this article - [Claude Managed Agents, which entered public beta April 8, 2026](https://www.tokenpost.com/news/technology/29146)