Anthropic's agent-memory playbook: progressive disclosure + dreaming

Four memory layers and an idle consolidation pass that lifts CL-Bench rewards far past stateless; plus a 14B video+audio model that runs on a single 5090.

Nowline NOV 8 12:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Four layers, and an agent that 'dreams'

    Anthropic's new write-up structures agent memory as four layers — transcripts, notes, representations, and a small always-injected context — read bottom-up only when needed, the same progressive-disclosure trick Claude Code uses to load just the first 25KB of MEMORY.md as an index. An idle “dreaming” pass that reconciles notes against the transcript took CL-Bench database reward from 0.655 to 0.838 for Opus 5.5, and the memory tool plus a dreaming cookbook are documented on the Claude Platform — so this is a pattern you can wire into your own agents, not just a paper.

  • Text-to-video with sound, on a single 5090

    Video Rebirth shipped HyperFlow H3 Flash, a 13.7B pruned-and-INT8 distill of MiniMax-H3 that makes 5–15s of 1344×768 video with synced 32kHz stereo audio in 4–8 steps — no classifier-free guidance. A 9GB Q4 GGUF runs on 24–32GB consumer GPUs (about 92s per 5s clip on an RTX 5090) with a Python CLI and ComfyUI nodes; the code is Apache-2.0, but the weights inherit MiniMax's community license that bars EU/UK/South Korea/US use without authorization — check that before you ship anything on it.