Tech • IA • Crypto
Memory and Dreaming for Self-Learning Agents
9/10TL;DR
Anthropic has introduced new memory and “dreaming” systems designed to help AI agents learn across tasks, improving performance over time and at scale.
KEY POINTS
Agents Are Expanding in Capability and Duration
AI agents are now capable of handling increasingly complex tasks over extended periods, with research indicating task duration capacity doubles roughly every seven months. This rapid progress has exposed a key limitation: maintaining useful context across long-running or repeated tasks remains difficult. Without persistent learning, agents effectively restart from scratch each time.
Memory Enables Continuous Learning
The new memory system allows agents to retain and reuse knowledge from previous tasks, improving outcomes over time. Instead of isolated task performance, agents can learn from prior mistakes, reuse successful strategies, and share insights across environments. This creates a compounding effect where each task benefits from accumulated experience.
File-Based Design Aligns With Model Strengths
Memory is structured as a file system, leveraging models’ strong ability to navigate, read, and modify files using familiar tools. This design minimizes friction and allows models to organize information naturally. The approach emphasizes flexibility, enabling agents to decide what information matters and how to store it effectively.
Multi-Agent Collaboration Through Shared Memory
The system supports shared memory across multiple agents, enabling collaboration within and across environments. Different access levels, such as read-only organizational memory and task-specific read-write stores, create a scalable hierarchy. This allows teams of agents to operate with a shared understanding while maintaining control over updates.
Enterprise Controls and Observability
Built-in safeguards include version control, audit trails, and attribution, allowing developers to track how memory evolves and which agent contributed specific changes. A standalone API enables integration across environments, with features like exports and redactions to meet enterprise requirements.
Measured Performance Gains in Production
Early adopters report significant improvements. Rakuten achieved a 97% reduction in first-pass errors, while Wise Docs reduced recurring issues in document verification workflows. These gains highlight the practical benefits of persistent memory in real-world deployments.
Limits of Local Memory Optimization
As systems scaled, issues emerged: duplicated knowledge, fragmented insights, and agents independently learning the same lessons. Memory updates were often locally optimal but lacked global coordination, limiting overall efficiency.
“Dreaming” Introduces Global Optimization
The new dreaming process acts as a feedback loop that analyzes activity across agents and sessions. It identifies patterns, recurring mistakes, and inefficiencies, then reorganizes and improves stored memory. This process runs independently from active tasks, ensuring no added latency.
Demonstrated Gains From Dreaming
Early results show strong impact, with Harvey reporting a sixfold increase in completion rates on legal benchmarks. By synthesizing insights across multiple runs, dreaming enables system-wide learning rather than isolated improvements.
Decoupled Architecture Enhances Performance
Dreaming operates خارج the main execution loop, allowing agents to focus on task completion while optimization happens separately. It can be triggered on schedules or events and processes multiple sessions in parallel, improving scalability without slowing active workflows.
Real-World Application in Incident Response
In one deployment, agents handling system alerts used shared memory to coordinate responses. When one agent identified a fix in progress, subsequent agents accessed that information and adjusted their actions accordingly. Dreaming later identified recurring patterns, such as alerts following CPU spikes, and updated memory to improve future triage decisions.
Toward Organizational-Scale Knowledge Systems
Combined, memory and dreaming create a continuously improving knowledge layer that spans agents and tasks. Memory captures experience, while dreaming refines and organizes it, effectively raising the baseline performance for all agents over time.
CONCLUSION
Anthropic’s memory and dreaming systems mark a shift toward AI agents that learn cumulatively and collaboratively, laying the groundwork for scalable, organization-wide intelligence.