Editorial illustration for Anthropic says Claude's agent memory requires forgetting, not just storing
Claude's Agent Memory Needs Forgetting, Not Storage
Anthropic's research team ran Claude through thousands of steps of Pokémon Red and found something that cuts against how most people think about AI memory: more storage isn't the fix for an agent that forgets everything between sessions. The company published a research post describing how it built note-taking into Claude so the model can write files to a folder, close out a session, and pick up roughly where it left off days later without the underlying weights changing at all.
The problem Anthropic ran into first was volume. Give an agent the ability to save notes, and it tends to save everything, because in the moment it has no idea what will matter later. That instinct produces archives nobody, including the model itself, can use.
Anthropic's tests point to a different approach: treating memory as a selection problem rather than a capacity problem, where deciding what to leave out matters more than how much gets kept. The company also built a process it calls "Dreaming," where Claude reviews its own past sessions during idle time to catch mistakes and clean up what it wrote.
Anthropic studied how Claude retains knowledge across sessions. Its tests show that selective notes outperform exhaustive archives and that having Claude review its own memories after the fact can correct errors.
Why this matters
For anyone building agents that run longer than a single session, Anthropic's findings push against the instinct to log everything "just in case." We've seen plenty of teams treat memory as a storage problem, bolt on bigger context windows or vector databases, and assume retrieval quality follows naturally. Anthropic's tests suggest the opposite: stronger models do better when given latitude to decide what's worth keeping, while rigid rules like "only record mistakes" actually hurt them. That's a useful check on over-engineering memory systems for frontier models.
The "Dreaming" step, reviewing transcripts during idle time to extract reusable strategy, also hints at where agent infrastructure is heading: less passive logging, more active curation that happens after the fact. Worth noting this benefits newer, more capable models more than older ones, which still need explicit guardrails. If that pattern holds elsewhere, it complicates one-size-fits-all memory architectures and means teams should be testing instruction rigidity against model capability, not assuming tighter rules are always safer.
Common Questions Answered
How does Anthropic's note-taking approach enable Claude to retain memory across sessions?
Anthropic built a note-taking system into Claude that allows the model to write files to a folder, close out a session, and resume work days later without changing the underlying weights. This approach enables Claude to pick up roughly where it left off by reviewing its own written notes from previous sessions, rather than relying solely on stored context.
Why does Anthropic say selective notes outperform exhaustive memory archives for agent memory?
Anthropic's research found that having Claude selectively decide what information is worth keeping produces better results than storing everything indiscriminately. The company's testing suggests that stronger models perform better when given the latitude to determine what's worth retaining, rather than forcing rigid rules that capture all data.
What did Anthropic discover about Claude's memory through its Pokémon Red experiments?
Anthropic ran Claude through thousands of steps of Pokémon Red and discovered that more storage isn't the solution for agents that forget everything between sessions. The research revealed that the key to agent memory is knowing what to forget, not how much to store, challenging conventional assumptions about memory capacity.
How can Claude correct errors in its own memories according to Anthropic's findings?
Anthropic's research shows that having Claude review its own memories after the fact can correct errors that may have occurred during initial note-taking. This self-review process allows the model to refine and validate the information it has stored across sessions.
Why should teams building longer-running agents reconsider treating memory as purely a storage problem?
Anthropic's findings suggest that bolting on bigger context windows or vector databases doesn't automatically improve retrieval quality, contrary to common assumptions. Instead, the research indicates that allowing models latitude to decide what's worth keeping produces superior results compared to rigid logging rules or exhaustive data capture strategies.
Further Reading
- The components of agent memory - Anthropic
- Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads - arXiv
- Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Open Challenges - arXiv
- Novel Memory Forgetting Techniques for Autonomous AI Agents - arXiv
- The Horizon Gap: Planning, Memory, Execution, Training, and Evaluation for Long-Horizon LLM Agents - arXiv