Tech • IA • Crypto
Agents that remember
9/10TL;DR
Anthropic has introduced persistent agent memory stores and an asynchronous “dreaming” system to enable AI agents to retain, refine, and reuse knowledge across sessions.
KEY POINTS
Limits of Stateless Agents
Most AI agent workflows operate in isolated sessions, meaning no information is retained once a session ends. This prevents continuity, forcing agents to relearn context repeatedly and limiting usefulness in real-world, multi-step tasks. As a result, even simple follow-up queries fail if they rely on prior interactions.
Introduction of Memory Stores
Memory stores act as persistent, file system-like storage layers attached to agent sessions. They allow agents to read and write data across sessions, effectively enabling long-term memory. Developers can create multiple stores scoped by user, workspace, or application, giving flexibility in how memory is organized.
File System-Based Architecture
The system mounts memory stores as accessible file systems within agent environments. Agents can search, read, and modify files using tools such as keyword search and shell commands. This structure allows complex data organization and retrieval, significantly improving agent capability compared to transient context windows.
Session Integration and Access Control
Memory stores are attached at the session level via API parameters, including configurable prompts to guide what information should be stored. Access can be set to read-write or read-only, enabling tighter control over whether agents can modify stored knowledge or only consume it.
Demonstrated Impact on Recall
When memory is enabled, agents can successfully retrieve and use information provided in previous sessions. In tests, agents without memory failed to answer follow-up questions, while those with memory accurately recalled stored details, demonstrating immediate practical benefits.
Manual and Programmatic Management
Developers can inspect, edit, and version memory files through both CLI and console interfaces. Each update creates a new version, allowing traceability and correction of errors. Human oversight is supported through direct editing and structured browsing of stored data.
The Problem of Memory Growth
Continuous writing to memory stores introduces new challenges, including unbounded growth, duplication, and outdated information. Without intervention, memory quality degrades over time, reducing retrieval efficiency and increasing computational cost.
“Dreaming” as Memory Optimization
The dreaming feature addresses these issues through asynchronous background processing. It analyzes stored memories and past session transcripts to consolidate data, remove duplicates, fact-check content, and enrich entries with additional details such as metadata and timelines.
Multi-Agent Processing Architecture
Dreaming operates as a multi-agent system, where an orchestrator assigns sub-agents to review individual session transcripts. This design ensures comprehensive coverage, enabling the system to process large volumes of historical data while maintaining structured outputs.
Non-Destructive Memory Refinement
Dreaming does not overwrite the original memory store. Instead, it creates a new, improved version, preserving the original data. This approach allows safe iteration and comparison, with developers able to adopt refined memory stores only after validation.
Enhanced Retrieval via Indexing
Output memory stores often include structured indexes and reorganized files, allowing faster and more precise information retrieval. Agents can consult these indexes to locate relevant data efficiently rather than performing broad searches.
Performance and Cost Considerations
Dreaming jobs can take minutes to hours depending on data size, but run asynchronously to avoid interrupting active workflows. Token usage is high by design, though caching can reach up to 95% efficiency, and model selection or scheduling can help reduce costs.
CONCLUSION
Persistent memory and asynchronous refinement mark a shift toward more autonomous, context-aware AI agents capable of accumulating and improving knowledge over time.