Anthropic reports accuracy and token gains from agent-built memory wikis

Anthropic reported that Claude answered questions about internal company documents more accurately and used fewer tokens when search was supplemented with a wiki built by agents. In its evaluation of 25 questions, search alone scored 53%; combining search with a growing wiki raised the aggregate score from 54% to 61% and saved about 20,000 tokens per question. The agents built linked pages during background sessions, checking claims against source documents and revising earlier notes. The corpus contained about 350,000 files, roughly 420M tokens. The gains varied by task. A legal benchmark task showed no clear improvement over search, while another produced a wiki larger than the underlying documents.