At its Code with Claude developers' conference, Anthropic has introduced a new feature called 'dreaming' to its Claude Managed Agents. Dreaming is a process that goes over recent events and identifies specific things worth storing in memory to inform future tasks and interactions.
The feature is currently in research preview and limited to Managed Agents on the Claude Platform. It's intended for situations where multiple agents work together on a task or project over several minutes or hours.
Anthropic describes dreaming as a scheduled process that reviews sessions and memory stores, curating specific memories. This is important because context windows are limited for large language models (LLMs), and important information can be lost over lengthy projects.
The company also mentioned compaction, a process used by many chat models to periodically analyze lengthy conversations and remove irrelevant information from the context window while keeping what's actually important for the ongoing conversation or task.
Dreaming is a periodically recurring process that analyzes past sessions and memory stores across agents, identifying and saving important patterns to memory for future use. Users will be able to choose between an automatic process or reviewing changes to memory directly.