Anthropic details agent memory architecture for Claude, including 'dreaming' process

Agent Memory Design Inspired by Human Cognition

Anthropic researchers have detailed the design of agent memory for their AI model, Claude. The approach draws inspiration from human memory processes, particularly the division of labor between immediate experience capture and long-term knowledge consolidation. This framework aims to enable Claude to retain knowledge from interactions and assigned tasks across sessions.

The Role of 'Dreaming' in Context Consolidation

A key component of Claude's memory system is a process referred to as 'dreaming'. Similar to how the human hippocampus replays experiences during sleep, Claude uses idle time between interactions to reorganize its existing context. This offline computation allows the model to reflect on logs, correct errors, and generate new notes or representations, integrating experiences into its knowledge base.

Salient Event Tagging for Retrieval

To facilitate efficient memory retrieval, Claude's system incorporates a mechanism for tagging salient events. Analogous to how chemical signals mark important memories in the human brain for more frequent replay, Claude can identify and tag significant events within a session log as they occur. This tagging makes these events easier to locate and utilize in subsequent interactions.

Distinction from Other Learning Mechanisms

Anthropic clarifies that this agent memory specifically refers to knowledge gained from experience and retained between sessions. It is distinct from other ways models learn or manage context, such as changes to internal parameters during training (weight changes), knowledge acquired within a single conversation (in-context learning), or summaries generated by Claude to manage context window limitations (compaction).