Definition
Plain language
An AI memory system with two parts — a fast scribe that writes notes after every task, and a slow editor that rewrites batches of notes between sessions.
As stated in the literature
A two-timescale memory architecture for language agents that pairs a per-session writer with an offline consolidator trained via GRPO; the consolidator wholesale-rewrites memory regions and is rewarded by task success plus a counterfactual subset-masking utility signal.
Why it matters: It separates fast capture from slow curation, which matches how human memory actually consolidates and avoids the trap of memory stores that just grow without ever getting better.
For example, after each task Auto-Dreamer's writer logs notes, and overnight the consolidator rewrites groups of related notes into cleaner, more reusable summaries.
Heard on the show
“It's called "Auto-Dreamer: Learning Offline Memory Consolidation for Language Agents.”Episode 064 — When Agent Memory Stops Being a Database and Starts Being a Skill