The Components of Agent Memory

Agent memory is knowledge retained across sessions from interactions and tasks, distinct from model training, within-conversation learning, or summaries used to continue a task. It can range from searchable transcripts to Claude-written notes, organized representations such as wikis, and information automatically loaded into future sessions. The central idea is that memory should be treated as a prediction about what will be useful later: agents must identify durable facts, tested rules, strategies, and lessons rather than preserve every detail. A “dreaming” process can review logs during idle periods, correct mistakes, consolidate notes, and shape what future sessions receive.

Evaluations show that effective memory can substantially improve long-horizon performance, but its value depends on how models organize and use it. On a database task from CL-Bench, stronger Claude models used feedback to validate rules, document evidence, and warn about traps, raising scores markedly over stateless runs. In Pokémon Red, Sonnet 3.7 accumulated hundreds of files describing changing, often redundant state; maintaining them consumed effort and yielded one badge. Fable 5 kept a compact set of goals, plans, and tested strategies and earned four badges, while Opus 5.5 also used concise, actively updated notes to beat the game. General permission to use memory worked better for capable models than rigid instructions such as recording only failures. Older models benefited more from explicit guidance to consult notes and test explanations, suggesting that memory design should let capable agents judge salience while providing weaker ones with targeted support.