RESEARCH · RESEARCH · #3669
Anthropic: selective notes and post-session "Dreaming" help Claude retain useful agent memory
Anthropic tested how Claude can persist knowledge across sessions by writing notes and then reviewing transcripts in an offline process called "Dreaming." In benchmarks (including a Pokémon Red playthrough and a 350,000‑document wiki build for Stainless), selective recording of strategies and rules plus Dreaming-based correction outperformed exhaustive logging and reduced errors, while Dreaming was offered as a research preview for Claude Managed Agents (initially supporting Opus 4.7 and Sonnet 4.6).
KEY POINTS
- Anthropic tested how Claude can persist knowledge across sessions by writing notes and then reviewing transcripts in an offline process called "Dreaming." In benchmarks (including a Pokémon Red playthrough and a 350,000‑document wiki build for Stainless), selective recording of strategies and rules plus Dreaming-based correction outperformed exhaustive logging and reduced errors, while Dreaming was offered as a research preview for Claude Managed Agents (initially supporting Opus 4.7 and Sonnet 4.6).
- This work shows that selection and post-session review (Dreaming) can make persistent agent memory more accurate and efficient, addressing stale or misleading notes that hurt multi‑session agent behavior.
- Anthropic says the key to agent memory is knowing what to forget, not how much to store
WHY IT MATTERS
This work shows that selection and post-session review (Dreaming) can make persistent agent memory more accurate and efficient, addressing stale or misleading notes that hurt multi‑session agent behavior.