Anthropic studies cross-session memory and offline cleanup for agents
What happened
On November 7, Hacker News' front page covered Anthropic's research on cross-session memory for Claude agents. Anthropic ran several experiments on what memory is made of and what it does, and studied a "dreaming" method that checks logs offline, corrects memories and builds linked knowledge pages. In a document library of about 350,000 files and roughly 420 million tokens, search plus wiki scored 54% and rose to 61% as the wiki grew; search alone scored 53%. Combining search and wiki also saved about 20,000 tokens per question.
Written by AI from the coverage · updated 17 hours ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageThe Components of Agent Memory
Anthropic 通过多项实验研究 Claude 的跨会话记忆，探索用 dreaming 离线核查日志、纠正记忆并构建关联知识页面。针对约 350,000 个文件、约 420M tokens 的文档库，结合搜索与 wiki 的评分随 wiki 扩展从 54% 升至 61%，仅搜索为 53%，每题还节省约 20,000 tokens。
Heat over time
Heat now 6·Comparable peak 10（Nov 7 16:00）·Comparable change over 24 hours –
The trend only compares accounts observed without gaps, so its range may be smaller than the current heat. Hover or tap the chart for each hour; the left and right arrow keys step through it.