Anthropic’s engineering team dropped a bombshell on November 6, 2026, revealing that strict instructions on what an agent should remember actually degrade performance. The study, titled "The components of agent memory," demonstrates that while persistent memory directories significantly boost benchmark scores for models like Claude Opus 4.7, imposing rigid rules—such as recording only failures—strips away much of that gain. This research isn't a new product launch but a critical deep dive into how long-lived agents should manage their internal state, directly impacting developers using Claude Code and Managed Agents.

The Benchmark Reality Check

The core experiment utilized the database task from CL-Bench 1.0, where agents answer questions across fresh sessions using a persistent memory directory. The results were stark: Claude Opus 4.7’s mean reward jumped from 0.383 (stateless) to 0.700 with memory. However, the outlier Sonnet 4.6 failed to benefit initially, scoring only 0.364, because it often ignored its own notes. When prompted to explicitly read and test theories from memory, its score nearly doubled to 0.641. This highlights a critical failure mode: a memory directory the model never opens is just wasted disk space.

Less Control, More Intelligence

The most counter-intuitive finding concerns steering. Developers instinctively want to control agent behavior with rules like "only record what went wrong." Anthropic found that for stronger models like Opus 5 and 4.7, this backfired. The report states, "it helps to tell Claude what memory is for, but not how to use it." Giving frontier models the freedom to judge salience proved more effective than fixed rules. In a Pokémon Red simulation, Sonnet 3.7 wrote 413 files and got one badge, while Fable 5 wrote nine files and earned four. Fewer, better-chosen notes correlated with higher progress.

The Four-Layer Hierarchy and Dreaming

Anthropic describes memory as a four-layer hierarchy: transcripts, memory notes, representations (like wikis), and injected context. The crucial warning is that the top layer—injected context—loads into every session and consumes valuable tokens. An incorrect note here propagates errors indefinitely. To mitigate this, Anthropic introduced "dreaming," an offline process where the agent reflects on logs to fix mistakes. In the wiki experiment, allowing Claude to build a structured knowledge base over 420M tokens raised graded criteria scores from 13% to 48%. This approach, combined with search, saved approximately 20,000 tokens per question compared to search alone.

Key Takeaways

  • Stop hard-coding memory rules; tell the agent the purpose, not the method.
  • Keep the injected context layer small; Claude Code caps this at 200 lines or 25KB.
  • Verify your model actually reads its memory files; older models may ignore them.
  • Use 'dreaming' to audit and correct notes against original transcripts during idle time.
  • For retrieval tasks, structured wikis built by the agent outperform raw search chunks.

The Bottom Line

If you are micromanaging your agent’s memory, you are likely throttling its intelligence. The era of hardcoded 'save this, ignore that' rules is over; let the model decide what matters, then use offline 'dreaming' to clean up the mess.