Here is another interesting read.
Anthropic explained how it organizes the memory of Claude's agents between sessions. The agent writes notes while working, and between the meeting you can run "dream": check notes with transcripts, correct errors, build a wiki and select the main thing for a small layer, which is loaded at the beginning of the next session.
In a sequence of my red poké (yes, agents play a Gameboy game), the Sonata 3.7 captured everything: 413 files in 225 folders and 1 icon. Example 5 was content with 9 files plus global.md and took 4 icons. Opus 5 agents built a wiki using internal stainless steel documents (about 420 million tokens), and over 128 dreaming sessions, the proportion of criteria fulfilled in responses to it increased from 13% to 48%.
Conclusion for your agents: In these tests, it helped strong models to explain why memory is needed and leave them to choose the facts. The strict "record only errors" rule cut the memory gain for Opus 5 by almost 90%, and for Opus 4.7 by about half.