Anthropic has published research on the memory architecture of AI agents

The company Anthropic has presented a conceptual description of a memory architecture for autonomous AI agents. The document systematizes approaches to organizing accumulated experience during system operation.

According to the research, agent memory is divided into short-term session logs and long-term structured notes. To consolidate, clean, and organize this information, the developers propose using a "sleep-time compute" mechanism.

The proposed scheme is intended to improve the efficiency of AI agents when performing long-term tasks that require maintaining context between individual user interaction sessions.