Anthropic lets Claude "dream" to organize memories: 350,000 files, Q&A score rises from 53% to 61%

Beating AI News Flash: Anthropic has published a study on long-term memory for Agents, detailing Claude's "Dreaming" mechanism. After continuously executing tasks, an Agent accumulates a large amount of conversations, operation logs, and notes. However, this information may be outdated, mutually contradictory, and may not be findable when truly needed.

Anthropic divides memory into four layers: raw records, curated notes, an interconnected Wiki knowledge base, and a brief index automatically loaded at each startup. Claude can jot down important information while executing tasks. "Dreaming" reviews historical records between tasks, corrects errors, deletes outdated content, and distills new experience from multiple tasks. At the next startup, Claude only needs to load the brief index and then look up relevant memories on demand, without having to re-read all history.

Anthropic conducted experiments using approximately 350,000 internal documents from Stainless, totaling 420 million tokens. Researchers had Opus 5 go through 128 rounds of sessions, continuously reading documents, building a Wiki, merging duplicate content, and verifying facts.

On 25 questions about these documents, using a search tool alone scored 53%, and adding a progressively refined Wiki raised it to 61%. Answering each question also saved an average of approximately 20,000 tokens. The Wiki's advantage is especially pronounced for questions that require synthesizing multiple contracts to answer; but if the answer is in a single document, ordinary search is already sufficient.

Anthropic also had Claude play "Pokémon Red." The older Sonnet 3.7 recorded 413 memory files and ultimately earned only one badge. The newer Fable 5 kept only 10 files but earned four badges. The newer model is better at distinguishing strategies worth saving from game states that can be re-checked at any time.