The Components of Agent Memory

160 points · 96 comments on HN · read original →

Points and comments are a snapshot, not live.

Anthropic describes agent memory using dreaming, hierarchies, and salience judgments inspired by human cognition.

Anthropic's post details how Claude agents can use memory inspired by human hippocampal replay. Key components include memory notes (written in-session based on predicted future relevance), dreaming (offline reflection on logs to correct and reorganize memories), and hierarchical promotion (floating key notes into context for new sessions). Evaluations on CL-Bench, Pokémon Red, and legal document analysis show that stronger models (Claude Opus 5, Sonnet 5, Fable 5) judge salience better than weaker ones, using fewer but more strategic notes. Dreaming helped Claude avoid looping in a Pokémon cave and improved wiki-building for document search from 13% to 48% over 128 sessions. Giving models freedom to decide what to remember outperformed fixed rules like "record only failures."

What commenters are saying

Top commenter expresses regret over having similar ideas ~2.5 years ago but not prototyping them, comparing it to missing Apple stock. They note Opus 5.5, Sonnet 5.5, and Opus 5 tied on the basic memory test, suggesting a plateau. They praise Figure 4's depiction of a "Mythos step-change" in agentic engineering. Another commenter offers empathy and advises sharing ideas freely to avoid feeling left behind. A reply counters that ideas are cheap; execution is rare. The thread splits between those validating the emotional struggle of unfulfilled ideation and those emphasizing execution over originality.