This article explores the design of memory systems for agents. A pertinent question alluded to is: how can an actor decide what to remember now when it doesn’t know what’s next? The ideas contain much anthropoligizing of Claude (comparing memory systems to the human brain). While I think there are problems with constantly seeking to understand agents through an anthropological lens, many of the connections are admittedly interesting. I can sympathize with the stated tradeoff where taking good notes while working on something can be distracting, but perfecting the notes later is resource-intensive. I appreciated the writing style that includes references to the humanities. A bit of Kierkegaard and Borges enriches almost any experience!

The memory system described contains a hierarchy (from top to bottom):

  • Injected context (smallest)
  • Memory representations (e.g. wikis)
  • Memory Notes
  • Transcripts (event logs)

“Dreaming” is then a process that moves information up the hierarchy while verifying against the layers below.

Observations

  • Opus 5.5 and Sonnet 5.5 had very similar scores on CL-Bench.
  • There is explicit mention of Claude being evaluated on its ability to play Pokemon Red. This connects to rumors that the advances gained in current frontier models largely come from reinforcement learning via playing games.
  • The better the model, the more selective it was in recording memories.
  • Giving explicit instructions for how to use memory helped older models much more than newer ones.
  • Providing a search tool in addition to raw file access to the memory wiki reduced token usage and improved eval scores.

Sources referenced: