How Claude manages memory

The video explains how to manage Claude's memory using claude.md files, which act as context for the stateless LLM. Claude loads these files to understand the conversation's history and project-specific instructions. Unlike other coding agents that standardize on agents.md, Claude uses claude.md, potentially requiring symlinking.

The speaker describes a tiered, hierarchical structure for claude.md files:

• Organization-wide: Managed by IT/DevOps.
• Project memory: claude.md files within the project, committed to Git, providing team-shared instructions like coding standards and common workflows.
• User-specific: claude.local.md, which is .git ignored and contains individual user preferences.

Auto-memory, another type of memory, infers things from conversations. It can be turned off and is stored in a specific directory structure under the system root. The speaker emphasizes that all claude.md files in the directory hierarchy are loaded at launch, impacting the context window.

The speaker shares a real-world example of a claude.md file filled with project-specific directions, emphasizing the importance of providing very specific instructions rather than general ones. He highlights examples such as coding standards, testing procedures, and platform-specific gotchas.

Claude can initialize a claude.md file with code-based documentation using the "Init" command. The speaker advises reviewing and adjusting the generated content to align with project specifics. Updating claude.md after errors is recommended, adding specific rules and guardrails based on past experiences to improve Claude's performance over time. Instead of dismissing AI coding assistance after an error, it is better to use that incident to refine the rules and systems surrounding the AI to ensure code quality.