AI BriefingThe components of agent memory
·2026.11.06 09:00
Key point
The study defines agent memory as persistent, experience-based knowledge and introduces 'dreaming' to correct errors, with benchmarks showing significant performance gains for newer Claude models.
Details
- Anthropic researchers define agent memory as experience-based knowledge maintained across sessions, distinct from model weights or in-context learning, and introduce 'dreaming' to reorganize context during idle time.
- On CL-Bench 1.0, memory access improved average rewards significantly: Claude Opus 5.5 rose from 0.655 to 0.838, and Claude Sonnet 5.5 from 0.644 to 0.833, while older models like Sonnet 4.6 saw minimal gains (0.330 to 0.364).
- In Pokémon Red gameplay, Claude Fable 5 achieved 4 badges by recording only strategic rules (10 files), whereas Claude Sonnet 3.7 recorded indiscriminately (413 files) and achieved only 1 badge.
- 'Dreaming' corrects memory errors by cross-referencing transcripts; in a Pokémon test, corrected memory allowed reaching Cerulean City at action 1,935, while uncorrected memory resulted in failure.
- Building a 'Wiki' representation via dreaming improved performance on a 25-question Stainless corpus evaluation. Using the wiki alone increased scores from 13% to 48% as sessions scaled from 2 to 128. Combining the wiki with a search tool raised the score from 54% to 61%, outperforming search alone (53%) and saving approximately 20,000 tokens per question.
- Steering strategies show that allowing frontier models to judge salience outperforms fixed rules; for older models like Sonnet 4.6, explicitly instructing memory reading nearly doubled benchmark scores (0.330 to 0.641).
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.