Anthropic has revealed details of a new memory system for its artificial intelligence (AI) assistant Claude, introducing a background mechanism called ‘Dreaming’ that lets the model review past interactions, correct inaccurate information, and organize knowledge for future use.
The mechanism works during idle periods, replaying previous interaction records to identify errors in stored memories and restructure information so it can be retrieved more efficiently in subsequent sessions.
The development addresses a key challenge in AI agents: retaining useful information across multiple conversations without accumulating excessive or outdated data.
According to Anthropic, agent memory lets Claude retain knowledge from user interactions, assigned tasks, and experiences across sessions. The company distinguishes this from model training, in-context learning within a single conversation, and compaction, which compresses earlier conversation turns into summaries.
Dreaming aims to reduce the need for AI models to repeatedly process entire interaction histories, which can increase token consumption and context length. Instead, it organizes accumulated information into a smaller, more accessible knowledge base.
Anthropic's experiments also found that letting stronger AI models decide what information deserves to be remembered can be more effective than imposing human-written memory rules.
Anthropic tested the approach using Pokémon Red, a game that requires long-term planning, navigation, and tracking of previous actions.
Claude Sonnet 3.7 created 413 memory files containing approximately 978,000 characters but earned only one badge. In comparison, Claude Fable 5 used just 10 memory files, containing around 43,000 characters, and earned four badges.
The experiment showed that retaining too much information can make it harder for AI agents to identify useful details and complete tasks.
Also Read: AI-Powered Vulnerability Detection: What GPT-6 Cyber Means for Developers
Anthropic also tested Dreaming using an internal document collection containing approximately 350,000 files and 420 million tokens.
After scaling the process from two sessions to 128 sessions, the rubric pass rate across 25 test questions increased from 13% to 48%. When the resulting wiki was combined with a search tool, the score rose further to 61%, while saving approximately 20,000 tokens per question.
Anthropic said the approach is not a complete solution, as excessive information and unnecessary detail can reduce the usefulness of generated knowledge.
The development highlights how competition in AI agents is increasingly shifting towards better memory organization, accuracy, and the ability to retain relevant information across sessions.