Introducing /remember: Teaching Agents to Learn from Experience

Key Takeaways

• Agents starting fresh each session lose valuable context accumulated during conversations
• The /remember slash command injects a reflection prompt into the conversation, triggering the agent to analyze the thread and extract learnings
• Persistent storage uses the file system: agents.md for preferences and user context, plus skill files for reusable workflows
• Agents can update memory organically during conversation when users say "always do this" or "this is a good workflow"
• Users retain full control—if /remember doesn't behave as expected, direct instructions override automated extraction

The /remember Flow

flowchart LR
    Conv[Conversation Thread] --> Remember["/remember command"]
    Remember --> Reflect[Agent Reflects]
    Reflect --> Decide{Skill or
Preference?}
    Decide -->|Preference| AgentsMD[(agents.md)]
    Decide -->|Workflow| Skills[(skills/)]
    AgentsMD & Skills --> Future[Future Sessions]

How It Works

The flow has three steps:

1. User invokes /remember — A prompt gets injected into the thread instructing the agent to review the conversation and capture valuable knowledge
2. Agent reflects — It analyzes the full conversation context, identifies best practices, and decides where to store each learning
3. Agent writes to persistent storage — Preferences go to agents.md, reusable workflows become skill files

Storage Structure

The file system serves as agent memory:

• agents.md — User preferences, ways of working, team context
• skills/ — Reusable workflows (e.g., "utility testing skill" with test-first development steps)

Notable Quotes

"Your thread is so good and you wish that you could bottle that up, bottle the information that you've shared in that thread and have the agent remember that."

"At the end of the day you have full control—if you decide 'I want this to be a skill' or 'I want this added to agents.md', you can just instruct the agent to do that."

Connections

• claude-code-continuous-learning-skill - Implements the same pattern for Claude Code: extracting reusable knowledge from debugging sessions and saving it as skills for future sessions
• agents-md-open-standard-for-ai-coding-agents - The agents.md file format this feature writes to, providing AI agents with project-specific guidance