Anthropic on agent memory: scaling 'dreaming' to 128 sessions lifts score from 13% to 48%
dl_weekly · x · 2026-11-18
Anthropic's engineering team published a deep dive into the components of agent memory, inspired by hippocampus-to-neocortex consolidation. Their 'dreaming' mechanism uses idle compute to reflect on logs, fix mistakes, and generate memory notes. Scaling dreaming from 2 to 128 sessions raised performance on 25 graded questions from 13% to 48%. The post covers memory notes, preemptive loading, and the evaluation suites they use.
More from coding & agent
- Just Screenshot the Bug: Skipping Copy-Paste When Tasking Coding Agents — fhinkel · 2026-11-18
- Free browser arcade golf game Fairway Fever built entirely with ChatGPT + Codex — ZabblesMarshmelon · 2026-11-18
- Dev builds Kasane Valley with Blender + three.js, showcasing 'GPT 6.2 Astra Ultra' workflow — OpenAIDevs · 2026-11-18
- 3 Weeks with GPT-6.2 Astra Internal Checkpoint: Less Lazy, Less Coding Slop — SIGKITTEN · 2026-11-18
- User files a Chase complaint addressed to the bank's AI agent, with a full image transcript — AaronBergman18 · 2026-11-18
- 10 people wrote 15,000 e2e tests in 2 weeks using this automation system — jdluk87 · 2026-11-18