Anthropic is moving beyond simple context windows with a new memory architecture for Claude agents, highlighted in a recent talk by Lamis, a member of technical staff on the applied AI team, at AI DevCon. The core problem remains persistent across LLM deployments: agents lack continuity, forcing users to re-explain project specifics every session. Lamis outlined a progression from static CLAUDE.md files to a dynamic, file-system-based memory approach that allows agents to read, write, and search their own knowledge bases during operation.

The Evolution of Agent Memory

The presentation detailed how Anthropic scaled memory management from simple markdown instruction files to more sophisticated systems. While CLAUDE.md files are effective for initial setup, they clog context windows as projects grow. Anthropic's current best practice involves treating memory as a file system, where agents use standard tools like bash and grep to search through markdown folders. This method decouples knowledge storage from the immediate context window, allowing for deeper detail retrieval without bloating the token count. Skills, or modular knowledge units, further optimize this by loading only titles initially and expanding content on demand, akin to pulling a specific book from a shelf only when needed.

Scaling Risks and Guardrails

Running a fleet of agents introduces significant concurrency and data integrity risks. Without proper safeguards, multiple agents can corrupt shared memory files simultaneously, or one agent can inject erroneous information that propagates across the entire system. Lamis emphasized four critical guardrails: versioning to track changes and enable rollbacks, concurrency controls using hashes to prevent write conflicts, strict permissions separating read-only company-wide knowledge from agent-specific scratchpads, and portability to ensure memory formats remain usable across different tools. These measures are essential for maintaining truth in a multi-agent environment where stale or malicious data can degrade performance.

Introducing 'Dreaming' for Background Optimization

The most notable addition to this framework is 'dreaming,' a background batch process designed to optimize memory outside of active user sessions. During live interactions, agents focus on task completion, often neglecting long-term memory hygiene. The dreaming process runs separately, analyzing past session transcripts—including tool calls and outcomes—to identify patterns, recurring errors, or outdated instructions. It proposes changes to the memory store, such as correcting a consistently failing tool configuration or updating a company-wide habit, which a human must then approve or reject. This separates the cognitive load of working from the meta-cognitive load of learning.

Practical Implementation and Community Tools

For developers looking to implement these concepts immediately, Anthropic’s Claude Managed Agents API includes built-in support for this memory and dreaming architecture, featuring versioning and hashing. Community efforts, such as the open-source 'claude-memory-kit' by Emmanuellsensai, provide a lightweight starting point using local markdown files and safety scripts. While the kit is not an official Anthropic product, it mirrors the recommended structure: a memory folder, skills, and a weekly 'dream' command that reviews recent sessions for improvements. The consensus is clear: start simple with markdown and search, but implement robust review cycles to keep agent knowledge accurate and relevant.

Key Takeaways

  • Memory is distinct from model intelligence; better context multiplies existing model capabilities.
  • File-system-based memory using markdown and grep/search tools is Anthropic's current best practice for scalability.
  • Concurrency, versioning, and permissions are mandatory guardrails when running multiple agents.
  • 'Dreaming' is a background process that analyzes past sessions to propose memory updates, reducing active session overhead.
  • Human review remains a critical step in the dreaming workflow to prevent the propagation of errors.

The Bottom Line

Anthropic is finally treating agent memory as a first-class engineering problem rather than an afterthought. The shift from static context to dynamic, file-based memory with background optimization is a necessary evolution for reliable production agents. If you are still pasting context manually, you are leaving significant performance gains on the table.