Anthropic’s November 6 engineering post describes agent memory as retained context that can be read in later sessions, rather than changes to model weights. Offline reflection, called dreaming, can reorganize notes and check them against transcripts.

Task-specific evaluations

On one CL-Bench database task, Anthropic reports Opus 5.5 mean reward of 0.655 without memory and 0.838 with memory. Those figures are vendor results for that task, not an overall benchmark score or a new model release.

In a separate document-corpus experiment, search plus a growing wiki rose from 54% to 61% of graded criteria across 25 questions, compared with 53% using search alone. Other document tasks showed no clear wiki benefit.

Maintenance matters

Notes can drift, contain errors or grow larger than the original documents. Reflection consumes compute and can delay availability of updates. The post does not establish a new memory feature enabled for every Claude user.