Søren Kierkegaard observed life must be lived forward while understood backward, mirroring the challenge of building memory for AI agents. Human memory relies on the hippocampus capturing experiences in real-time, then consolidating key moments during offline periods like sleep through replay, integrating them into long-term neocortex knowledge.
Researchers at Letta and UC Berkeley propose "sleep-time compute" for agents, using idle time between interactions to reorganize context and refine memory by reflecting on logs — a process they call "dreaming," which has proven effective in enhancing Claude's memory retention.
"Agent memory" refers to knowledge Claude retains between sessions, stored as context it can read later. This excludes weight changes, in-context learning, and compaction. Like humans, Claude tags salient events during sessions or writes them to memory stores, but must predict what's important before knowing future needs.
For example, in the Continual Learning Bench 1.0 task, Claude writes verified facts as predictive rules for future use. When incorrect, it records failures, theorizes causes, and creates rules for improvement. Using persistent memory directories across sessions, different models showed varying success in leveraging this learning loop.
Source: Hacker News · Summarized by HeadlinesBriefing