Anthropic breaks agent memory into notes, salience and log processing between sessions
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
An agent's long-term memory requires more than a large context: the system must select useful experience, correct its notes and prepare them for future tasks.
In Anthropic's engineering analysis, agent memory means knowledge from past interactions and assignments retained as context available in later sessions, rather than changes to model weights or compression of the current dialogue. Claude can mark important events as they happen, write them to files and later use those notes. A separate mechanism called dreaming processes logs between interactions: it corrects mistakes, creates notes and new memory representations, and selects information to load into future sessions. Anthropic compares this with memory consolidation during sleep and uses Continual Learning Bench tasks to show how an agent can formulate a rule after an error and apply it to the next question.
Sources 1