Today for AI
HOT RADAR
agentHEAT 18.3°

AI CURATED HOT STORY · 11/8/2026

Anthropic Details Agent Memory Design, Adding a Dreaming Step to Correct Notes

1 reports archived1 independent sourcesupdated 11/8/2026, 11:21:33
Synthesis & Latest Updates
1 Sources Cross-Validated

Anthropic's engineering team detailed a four-layer memory architecture for Claude spanning raw session logs, self-written notes, structured wikis, and auto-injected context, plus a post-session 'dreaming' step that revisits transcripts to fix and reorganize notes. Pokemon and database-QA tests show memory quality, not volume, drives performance, with Opus 5 the first model to form a complete learning loop.

LATEST/Anthropic's engineering team detailed a four-layer memory architecture for Claude spanning raw session logs, self-written notes, structured wikis, and auto-injected context, plus a post-session 'dreaming' step that revisits transcripts to fix and reorganize notes. Pokemon and database-QA tests show memory quality, not volume, drives performance, with Opus 5 the first model to form a complete learning loop.

Hourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first
  1. 宝玉 (@dotey)T2·78 pts

    Anthropic Details Agent Memory Design, Adding a Dreaming Step to Correct Notes

    • Memory has four layers: raw transcripts, self-written notes, structured wikis, and auto-injected context, with 'dreaming' reconciling and promoting content upward
    • In the Pokemon test, Sonnet 3.7 wrote 413 memory files for 1 badge while Fable 5 wrote 10 files for 4 badges, showing selectivity beats volume
    • On database QA, Opus 5 rose from 0.564 to 0.828 with memory, the first model to close a full learning loop